Regina Barzilay: Deep Learning for Cancer Diagnosis and Treatment
AI 与机器学习技术与编程心理与人性生物与进化音乐与艺术
📋 章节目录
暂无章节信息
🔑 关键词
datamachinecancerdonlearningsciencedoingareamitbookthinkingcomputerinterestingunderstandinglanguagetryinghumanmoleculeshospitaldrugs
💬 精彩语录
暂无语录
🎙️ 完整对话（1621 条）
Lex Fridman (00:00.000)
The following is a conversation with Regina Barzilay.
以下是与雷吉娜·巴尔齐莱的对话。
Lex Fridman (00:03.220)
She's a professor at MIT and a world class researcher
她是麻省理工学院的教授和世界级的研究员
Lex Fridman (00:06.700)
in natural language processing
在自然语言处理中
Lex Fridman (00:08.340)
and applications of deep learning to chemistry and oncology
以及深度学习在化学和肿瘤学中的应用
Lex Fridman (00:12.460)
or the use of deep learning for early diagnosis,
或利用深度学习进行早期诊断，
Regina Barzilay (00:15.340)
prevention and treatment of cancer.
预防和治疗癌症。
Lex Fridman (00:18.300)
She has also been recognized for teaching
她的教学能力也得到认可
Regina Barzilay (00:21.020)
of several successful AI related courses at MIT,
麻省理工学院几门成功的人工智能相关课程，
Lex Fridman (00:24.700)
including the popular Introduction
包括流行的介绍
Regina Barzilay (00:26.840)
to Machine Learning course.
机器学习课程。
Lex Fridman (00:28.920)
This is the Artificial Intelligence podcast.
这是人工智能播客。
Regina Barzilay (00:32.160)
If you enjoy it, subscribe on YouTube,
如果您喜欢，请在 YouTube 上订阅，
Lex Fridman (00:34.560)
give it five stars on iTunes, support it on Patreon
在 iTunes 上给它五颗星，在 Patreon 上支持它
Regina Barzilay (00:37.840)
or simply connect with me on Twitter
或者直接在 Twitter 上与我联系
Lex Fridman (00:39.840)
at Lex Friedman spelled F R I D M A N.
Lex Friedman 拼写为 F R I D M A N。
Lex Fridman (00:43.760)
And now here's my conversation with Regina Barzilay.
现在这是我与雷吉娜·巴尔齐莱的对话。
Lex Fridman (00:48.840)
In an interview you've mentioned
在一次采访中你提到过
Regina Barzilay (00:50.320)
that if there's one course you would take,
如果有一门课程你会选修
Lex Fridman (00:51.960)
it would be a literature course with a friend of yours
这将是与你的朋友一起上的文学课程
Regina Barzilay (00:54.600)
that a friend of yours teaches.
你的一个朋友教的。
Lex Fridman (00:56.360)
Just out of curiosity, because I couldn't find anything
Regina Barzilay (00:59.160)
on it, are there books or ideas that had profound impact
Lex Fridman (01:04.400)
on your life journey, books and ideas perhaps
Lex Fridman (01:07.200)
outside of computer science and the technical fields?
Lex Fridman (01:11.780)
I think because I'm spending a lot of my time at MIT
Lex Fridman (01:14.680)
and previously in other institutions where I was a student,
Lex Fridman (01:18.280)
I have limited ability to interact with people.
Lex Fridman (01:21.040)
So a lot of what I know about the world
Lex Fridman (01:22.640)
actually comes from books.
Lex Fridman (01:24.220)
And there were quite a number of books
Lex Fridman (01:27.240)
that had profound impact on me and how I view the world.
Regina Barzilay (01:31.380)
Let me just give you one example of such a book.
Lex Fridman (01:35.820)
I've maybe a year ago read a book
Regina Barzilay (01:39.660)
called The Emperor of All Melodies.
Lex Fridman (01:42.500)
It's a book about, it's kind of a history of science book
Regina Barzilay (01:45.740)
on how the treatments and drugs for cancer were developed.
Lex Fridman (01:50.740)
And that book, despite the fact that I am in the business
Regina Barzilay (01:54.580)
of science, really opened my eyes on how imprecise
Lex Fridman (01:59.780)
and imperfect the discovery process is
Lex Fridman (02:03.060)
and how imperfect our current solutions
Lex Fridman (02:06.980)
and what makes science succeed and be implemented.
Lex Fridman (02:11.060)
And sometimes it's actually not the strengths of the idea,
Lex Fridman (02:14.100)
but devotion of the person who wants to see it implemented.
Lex Fridman (02:17.420)
So this is one of the books that, you know,
Lex Fridman (02:19.780)
at least for the last year, quite changed the way
Regina Barzilay (02:22.300)
I'm thinking about scientific process
Lex Fridman (02:24.940)
just from the historical perspective
Lex Fridman (02:26.700)
and what do I need to do to make my ideas really implemented.
Lex Fridman (02:33.460)
Let me give you an example of a book
Regina Barzilay (02:36.060)
which is not kind of, which is a fiction book.
Lex Fridman (02:40.620)
It's a book called Americana.
Lex Fridman (02:44.420)
And this is a book about a young female student
Lex Fridman (02:48.780)
who comes from Africa to study in the United States.
Lex Fridman (02:53.260)
And it describes her past, you know, within her studies
Lex Fridman (02:57.740)
and her life transformation that, you know,
Regina Barzilay (03:02.020)
in a new country and kind of adaptation to a new culture.
Lex Fridman (03:06.540)
And when I read this book, I saw myself
Regina Barzilay (03:11.220)
in many different points of it,
Lex Fridman (03:13.540)
but it also kind of gave me the lens on different events.
Lex Fridman (03:20.140)
And some of it that I never actually paid attention.
Lex Fridman (03:22.060)
One of the funny stories in this book
Regina Barzilay (03:24.700)
is how she arrives to her new college
Lex Fridman (03:30.420)
and she starts speaking in English
Lex Fridman (03:32.900)
and she had this beautiful British accent
Lex Fridman (03:35.700)
because that's how she was educated in her country.
Regina Barzilay (03:39.860)
This is not my case.
Lex Fridman (03:40.980)
And then she notices that the person who talks to her,
Regina Barzilay (03:45.460)
you know, talks to her in a very funny way,
Lex Fridman (03:47.220)
in a very slow way.
Lex Fridman (03:48.340)
And she's thinking that this woman is disabled
Lex Fridman (03:51.460)
and she's also trying to kind of to accommodate her.
Lex Fridman (03:54.500)
And then after a while, when she finishes her discussion
Lex Fridman (03:56.700)
with this officer from her college,
Regina Barzilay (03:59.860)
she sees how she interacts with the other students,
Lex Fridman (04:02.100)
with American students.
Lex Fridman (04:03.020)
And she discovers that actually she talked to her this way
Lex Fridman (04:08.020)
because she saw that she doesn't understand English.
Lex Fridman (04:11.020)
And I thought, wow, this is a funny experience.
Lex Fridman (04:14.180)
And literally within few weeks,
Regina Barzilay (04:16.940)
I went to LA to a conference
Lex Fridman (04:20.820)
and I asked somebody in the airport,
Regina Barzilay (04:23.180)
you know, how to find like a cab or something.
Lex Fridman (04:25.580)
And then I noticed that this person is talking
Regina Barzilay (04:28.380)
in a very strange way.
Lex Fridman (04:29.220)
And my first thought was that this person
Regina Barzilay (04:31.100)
have some, you know, pronunciation issues or something.
Lex Fridman (04:34.500)
And I'm trying to talk very slowly to him
Lex Fridman (04:36.060)
and I was with another professor, Ernst Frankel.
Lex Fridman (04:38.580)
And he's like laughing because it's funny
Regina Barzilay (04:42.180)
that I don't get that the guy is talking in this way
Lex Fridman (04:44.860)
because he thinks that I cannot speak.
Lex Fridman (04:46.060)
So it was really kind of mirroring experience.
Lex Fridman (04:49.100)
And it led me think a lot about my own experiences
Regina Barzilay (04:53.300)
moving, you know, from different countries.
Lex Fridman (04:56.060)
So I think that books play a big role
Regina Barzilay (04:59.300)
in my understanding of the world.
Lex Fridman (05:01.780)
On the science question, you mentioned that
Regina Barzilay (05:06.420)
it made you discover that personalities of human beings
Lex Fridman (05:09.780)
are more important than perhaps ideas.
Lex Fridman (05:12.420)
Is that what I heard?
Lex Fridman (05:13.660)
It's not necessarily that they are more important
Regina Barzilay (05:15.740)
than ideas, but I think that ideas on their own
Lex Fridman (05:19.180)
are not sufficient.
Lex Fridman (05:20.460)
And many times, at least at the local horizon,
Lex Fridman (05:24.660)
it's the personalities and their devotion to their ideas
Regina Barzilay (05:29.140)
is really that locally changes the landscape.
Lex Fridman (05:32.980)
Now, if you're looking at AI, like let's say 30 years ago,
Regina Barzilay (05:37.500)
you know, dark ages of AI or whatever,
Lex Fridman (05:39.180)
what is symbolic times, you can use any word.
Regina Barzilay (05:42.420)
You know, there were some people,
Lex Fridman (05:44.660)
now we're looking at a lot of that work
Lex Fridman (05:46.620)
and we're kind of thinking this was not really
Lex Fridman (05:48.780)
maybe a relevant work, but you can see that some people
Regina Barzilay (05:52.220)
managed to take it and to make it so shiny
Lex Fridman (05:54.900)
and dominate the academic world
Lex Fridman (05:59.260)
and make it to be the standard.
Lex Fridman (06:02.380)
If you look at the area of natural language processing,
Regina Barzilay (06:06.420)
it is well known fact that the reason that statistics
Lex Fridman (06:09.140)
in NLP took such a long time to become mainstream
Regina Barzilay (06:13.980)
because there were quite a number of personalities
Lex Fridman (06:16.860)
which didn't believe in this idea
Lex Fridman (06:18.460)
and didn't stop research progress in this area.
Lex Fridman (06:22.060)
So I do not think that, you know,
Regina Barzilay (06:25.900)
kind of asymptotically maybe personalities matters,
Lex Fridman (06:28.940)
but I think locally it does make quite a bit of impact
Lex Fridman (06:33.940)
and it's generally, you know,
Lex Fridman (06:36.900)
speeds up the rate of adoption of the new ideas.
Regina Barzilay (06:41.340)
Yeah, and the other interesting question
Lex Fridman (06:43.500)
is in the early days of particular discipline,
Regina Barzilay (06:46.540)
I think you mentioned in that book
Lex Fridman (06:50.460)
is ultimately a book of cancer.
Regina Barzilay (06:52.340)
It's called The Emperor of All Melodies.
Lex Fridman (06:55.100)
Yeah, and those melodies included the trying to,
Lex Fridman (06:58.580)
the medicine, was it centered around?
Lex Fridman (07:00.740)
So it was actually centered on, you know,
Lex Fridman (07:04.900)
how people thought of curing cancer.
Lex Fridman (07:07.180)
Like for me, it was really a discovery how people,
Lex Fridman (07:10.660)
what was the science of chemistry behind drug development
Lex Fridman (07:14.140)
that it actually grew up out of dying,
Regina Barzilay (07:17.220)
like coloring industry that people
Lex Fridman (07:19.780)
who developed chemistry in 19th century in Germany
Lex Fridman (07:23.780)
and Britain to do, you know, the really new dyes.
Lex Fridman (07:28.140)
They looked at the molecules and identified
Regina Barzilay (07:30.180)
that they do certain things to cells.
Lex Fridman (07:32.140)
And from there, the process started.
Regina Barzilay (07:34.500)
And, you know, like historically saying,
Lex Fridman (07:35.740)
yeah, this is fascinating
Regina Barzilay (07:36.900)
that they managed to make the connection
Lex Fridman (07:38.700)
and look under the microscope and do all this discovery.
Lex Fridman (07:42.300)
But as you continue reading about it
Lex Fridman (07:44.340)
and you read about how chemotherapy drugs
Regina Barzilay (07:48.780)
which were developed in Boston,
Lex Fridman (07:50.500)
and some of them were developed.
Lex Fridman (07:52.500)
And Farber, Dr. Farber from Dana Farber,
Lex Fridman (07:57.500)
you know, how the experiments were done
Regina Barzilay (08:00.460)
that, you know, there was some miscalculation,
Lex Fridman (08:03.340)
let's put it this way.
Lex Fridman (08:04.540)
And they tried it on the patients and they just,
Lex Fridman (08:06.740)
and those were children with leukemia and they died.
Lex Fridman (08:09.980)
And then they tried another modification.
Lex Fridman (08:11.660)
You look at the process, how imperfect is this process?
Regina Barzilay (08:15.020)
And, you know, like, if we're again looking back
Lex Fridman (08:17.500)
like 60 years ago, 70 years ago,
Regina Barzilay (08:19.180)
you can kind of understand it.
Lex Fridman (08:20.780)
But some of the stories in this book
Regina Barzilay (08:23.020)
which were really shocking to me
Lex Fridman (08:24.620)
were really happening, you know, maybe decades ago.
Lex Fridman (08:27.980)
And we still don't have a vehicle
Lex Fridman (08:30.660)
to do it much more fast and effective and, you know,
Regina Barzilay (08:35.100)
scientific the way I'm thinking computer science scientific.
Lex Fridman (08:38.220)
So from the perspective of computer science,
Regina Barzilay (08:40.420)
you've gotten a chance to work the application to cancer
Lex Fridman (08:43.780)
and to medicine in general.
Regina Barzilay (08:44.860)
From a perspective of an engineer and a computer scientist,
Lex Fridman (08:48.420)
how far along are we from understanding the human body,
Regina Barzilay (08:51.780)
biology of being able to manipulate it
Lex Fridman (08:55.140)
in a way we can cure some of the maladies,
Lex Fridman (08:57.940)
some of the diseases?
Lex Fridman (08:59.740)
So this is very interesting question.
Lex Fridman (09:03.460)
And if you're thinking as a computer scientist
Lex Fridman (09:06.020)
about this problem, I think one of the reasons
Regina Barzilay (09:09.820)
that we succeeded in the areas
Lex Fridman (09:11.900)
we as a computer scientist succeeded
Regina Barzilay (09:13.980)
is because we don't have,
Lex Fridman (09:16.260)
we are not trying to understand in some ways.
Regina Barzilay (09:18.980)
Like if you're thinking about like eCommerce, Amazon,
Lex Fridman (09:22.260)
Amazon doesn't really understand you.
Lex Fridman (09:24.220)
And that's why it recommends you certain books
Lex Fridman (09:27.700)
or certain products, correct?
Regina Barzilay (09:30.660)
And, you know, traditionally when people
Lex Fridman (09:34.660)
were thinking about marketing, you know,
Regina Barzilay (09:36.380)
they divided the population to different kind of subgroups,
Lex Fridman (09:39.780)
identify the features of this subgroup
Lex Fridman (09:41.740)
and come up with a strategy
Lex Fridman (09:43.140)
which is specific to that subgroup.
Regina Barzilay (09:45.580)
If you're looking about recommendation system,
Lex Fridman (09:47.340)
they're not claiming that they're understanding somebody,
Regina Barzilay (09:50.580)
they're just managing to,
Lex Fridman (09:52.700)
from the patterns of your behavior
Regina Barzilay (09:54.780)
to recommend you a product.
Lex Fridman (09:57.540)
Now, if you look at the traditional biology,
Lex Fridman (09:59.580)
and obviously I wouldn't say that I
Lex Fridman (10:03.180)
at any way, you know, educated in this field,
Lex Fridman (10:06.180)
but you know what I see, there's really a lot of emphasis
Lex Fridman (10:09.300)
on mechanistic understanding.
Lex Fridman (10:10.660)
And it was very surprising to me
Lex Fridman (10:12.540)
coming from computer science,
Lex Fridman (10:13.820)
how much emphasis is on this understanding.
Lex Fridman (10:17.580)
And given the complexity of the system,
Regina Barzilay (10:20.740)
maybe the deterministic full understanding
Lex Fridman (10:23.220)
of this process is, you know, beyond our capacity.
Lex Fridman (10:27.380)
And the same ways in computer science
Lex Fridman (10:29.460)
when we're doing recognition, when you do recommendation
Lex Fridman (10:31.540)
and many other areas,
Lex Fridman (10:32.780)
it's just probabilistic matching process.
Lex Fridman (10:35.940)
And in some way, maybe in certain cases,
Lex Fridman (10:40.100)
we shouldn't even attempt to understand
Regina Barzilay (10:42.940)
or we can attempt to understand, but in parallel,
Lex Fridman (10:45.780)
we can actually do this kind of matchings
Regina Barzilay (10:48.060)
that would help us to find key role
Lex Fridman (10:51.060)
to do early diagnostics and so on.
Lex Fridman (10:54.100)
And I know that in these communities,
Lex Fridman (10:55.860)
it's really important to understand,
Lex Fridman (10:59.060)
but I'm sometimes wondering, you know,
Lex Fridman (11:00.700)
what exactly does it mean to understand here?
Regina Barzilay (11:02.940)
Well, there's stuff that works and,
Lex Fridman (11:05.500)
but that can be, like you said,
Regina Barzilay (11:07.620)
separate from this deep human desire
Lex Fridman (11:10.340)
to uncover the mysteries of the universe,
Regina Barzilay (11:12.700)
of science, of the way the body works,
Lex Fridman (11:16.140)
the way the mind works.
Regina Barzilay (11:17.620)
It's the dream of symbolic AI,
Lex Fridman (11:19.540)
of being able to reduce human knowledge into logic
Lex Fridman (11:25.220)
and be able to play with that logic
Lex Fridman (11:26.900)
in a way that's very explainable
Lex Fridman (11:28.700)
and understandable for us humans.
Lex Fridman (11:30.300)
I mean, that's a beautiful dream.
Lex Fridman (11:31.780)
So I understand it, but it seems that
Lex Fridman (11:34.860)
what seems to work today and we'll talk about it more
Regina Barzilay (11:37.900)
is as much as possible, reduce stuff into data,
Lex Fridman (11:40.780)
reduce whatever problem you're interested in to data
Lex Fridman (11:43.900)
and try to apply statistical methods,
Lex Fridman (11:47.060)
apply machine learning to that.
Regina Barzilay (11:49.100)
On a personal note,
Lex Fridman (11:51.140)
you were diagnosed with breast cancer in 2014.
Lex Fridman (11:55.380)
What did facing your mortality make you think about?
Lex Fridman (11:58.420)
How did it change you?
Regina Barzilay (12:00.260)
You know, this is a great question
Lex Fridman (12:01.860)
and I think that I was interviewed many times
Lex Fridman (12:03.820)
and nobody actually asked me this question.
Lex Fridman (12:05.740)
I think I was 43 at a time.
Lex Fridman (12:09.700)
And the first time I realized in my life that I may die
Lex Fridman (12:12.860)
and I never thought about it before.
Lex Fridman (12:14.460)
And there was a long time since you're diagnosed
Lex Fridman (12:17.260)
until you actually know what you have
Lex Fridman (12:18.580)
and how severe is your disease.
Lex Fridman (12:20.180)
For me, it was like maybe two and a half months.
Lex Fridman (12:23.500)
And I didn't know where I am during this time
Lex Fridman (12:28.340)
because I was getting different tests
Lex Fridman (12:30.660)
and one would say it's bad and I would say, no, it is not.
Lex Fridman (12:33.380)
So until I knew where I am,
Regina Barzilay (12:34.900)
I really was thinking about
Lex Fridman (12:36.300)
all these different possible outcomes.
Regina Barzilay (12:38.220)
Were you imagining the worst
Lex Fridman (12:39.700)
or were you trying to be optimistic or?
Regina Barzilay (12:41.940)
It would be really,
Lex Fridman (12:43.540)
I don't remember what was my thinking.
Regina Barzilay (12:47.340)
It was really a mixture with many components at the time
Lex Fridman (12:51.100)
speaking in our terms.
Lex Fridman (12:54.100)
And one thing that I remember,
Lex Fridman (12:59.340)
and every test comes and then you're saying,
Regina Barzilay (13:01.500)
oh, it could be this or it may not be this.
Lex Fridman (13:03.300)
And you're hopeful and then you're desperate.
Lex Fridman (13:04.700)
So it's like, there is a whole slew of emotions
Lex Fridman (13:07.660)
that goes through you.
Lex Fridman (13:09.820)
But what I remember is that when I came back to MIT,
Lex Fridman (13:15.100)
I was kind of going the whole time through the treatment
Regina Barzilay (13:17.780)
to MIT, but my brain was not really there.
Lex Fridman (13:19.780)
But when I came back, really finished my treatment
Lex Fridman (13:21.820)
and I was here teaching and everything,
Lex Fridman (13:24.900)
I look back at what my group was doing,
Lex Fridman (13:27.060)
what other groups was doing.
Lex Fridman (13:28.820)
And I saw these trivialities.
Regina Barzilay (13:30.820)
It's like people are building their careers
Lex Fridman (13:33.260)
on improving some parts around two or 3% or whatever.
Regina Barzilay (13:36.900)
I was, it's like, seriously,
Lex Fridman (13:38.380)
I did a work on how to decipher ugaritic,
Regina Barzilay (13:40.740)
like a language that nobody speak and whatever,
Lex Fridman (13:42.860)
like what is significance?
Regina Barzilay (13:46.140)
When all of a sudden, I walked out of MIT,
Lex Fridman (13:49.020)
which is when people really do care
Lex Fridman (13:51.860)
what happened to your ICLR paper,
Lex Fridman (13:54.500)
what is your next publication to ACL,
Regina Barzilay (13:57.900)
to the world where people, you see a lot of suffering
Lex Fridman (14:01.860)
that I'm kind of totally shielded on it on daily basis.
Lex Fridman (14:04.900)
And it's like the first time I've seen like real life
Lex Fridman (14:07.460)
and real suffering.
Lex Fridman (14:09.700)
And I was thinking, why are we trying to improve the parser
Lex Fridman (14:13.260)
or deal with trivialities when we have capacity
Lex Fridman (14:18.340)
to really make a change?
Lex Fridman (14:20.700)
And it was really challenging to me because on one hand,
Regina Barzilay (14:24.620)
I have my graduate students really want to do their papers
Lex Fridman (14:27.420)
and their work, and they want to continue to do
Lex Fridman (14:29.860)
what they were doing, which was great.
Lex Fridman (14:31.900)
And then it was me who really kind of reevaluated
Lex Fridman (14:36.300)
what is the importance.
Lex Fridman (14:37.460)
And also at that point, because I had to take some break,
Regina Barzilay (14:42.500)
I look back into like my years in science
Lex Fridman (14:47.740)
and I was thinking, like 10 years ago,
Regina Barzilay (14:50.460)
this was the biggest thing, I don't know, topic models.
Lex Fridman (14:52.940)
We have like millions of papers on topic models
Lex Fridman (14:55.340)
and variation of topics models.
Lex Fridman (14:56.500)
Now it's totally like irrelevant.
Lex Fridman (14:58.580)
And you start looking at this, what do you perceive
Lex Fridman (15:02.460)
as important at different point of time
Lex Fridman (15:04.500)
and how it fades over time.
Lex Fridman (15:08.900)
And since we have a limited time,
Regina Barzilay (15:12.980)
all of us have limited time on us,
Lex Fridman (15:14.900)
it's really important to prioritize things
Regina Barzilay (15:18.380)
that really matter to you, maybe matter to you
Lex Fridman (15:20.540)
at that particular point.
Lex Fridman (15:22.020)
But it's important to take some time
Lex Fridman (15:24.380)
and understand what matters to you,
Regina Barzilay (15:26.940)
which may not necessarily be the same
Lex Fridman (15:28.860)
as what matters to the rest of your scientific community
Lex Fridman (15:31.700)
and pursue that vision.
Lex Fridman (15:34.580)
So that moment, did it make you cognizant?
Regina Barzilay (15:38.460)
You mentioned suffering of just the general amount
Lex Fridman (15:42.500)
of suffering in the world.
Lex Fridman (15:44.340)
Is that what you're referring to?
Lex Fridman (15:45.620)
So as opposed to topic models
Lex Fridman (15:47.420)
and specific detailed problems in NLP,
Lex Fridman (15:50.780)
did you start to think about other people
Lex Fridman (15:54.460)
who have been diagnosed with cancer?
Lex Fridman (15:56.940)
Is that the way you started to see the world perhaps?
Regina Barzilay (16:00.020)
Oh, absolutely.
Lex Fridman (16:00.860)
And it actually creates, because like, for instance,
Regina Barzilay (16:04.980)
there is parts of the treatment
Lex Fridman (16:05.820)
where you need to go to the hospital every day
Lex Fridman (16:08.500)
and you see the community of people that you see
Lex Fridman (16:11.620)
and many of them are much worse than I was at a time.
Lex Fridman (16:16.100)
And you all of a sudden see it all.
Lex Fridman (16:20.460)
And people who are happier someday
Regina Barzilay (16:23.940)
just because they feel better.
Lex Fridman (16:25.300)
And for people who are in our normal realm,
Regina Barzilay (16:28.500)
you take it totally for granted that you feel well,
Lex Fridman (16:30.820)
that if you decide to go running, you can go running
Lex Fridman (16:32.940)
and you're pretty much free
Lex Fridman (16:35.900)
to do whatever you want with your body.
Regina Barzilay (16:37.620)
Like I saw like a community,
Lex Fridman (16:40.180)
my community became those people.
Lex Fridman (16:42.820)
And I remember one of my friends, Dina Katabi,
Lex Fridman (16:47.460)
took me to Prudential to buy me a gift for my birthday.
Lex Fridman (16:50.420)
And it was like the first time in months
Lex Fridman (16:52.340)
that I went to kind of to see other people.
Lex Fridman (16:54.980)
And I was like, wow, first of all, these people,
Lex Fridman (16:58.180)
they are happy and they're laughing
Lex Fridman (16:59.820)
and they're very different from these other my people.
Lex Fridman (17:02.620)
And second of thing, I think it's totally crazy.
Regina Barzilay (17:04.620)
They're like laughing and wasting their money
Lex Fridman (17:06.620)
on some stupid gifts.
Lex Fridman (17:08.420)
And they may die.
Lex Fridman (17:12.540)
They already may have cancer and they don't understand it.
Lex Fridman (17:15.940)
So you can really see how the mind changes
Lex Fridman (17:20.060)
that you can see that,
Regina Barzilay (17:22.340)
before that you can ask,
Lex Fridman (17:23.180)
didn't you know that you're gonna die?
Regina Barzilay (17:24.380)
Of course I knew, but it was a kind of a theoretical notion.
Lex Fridman (17:28.340)
It wasn't something which was concrete.
Lex Fridman (17:31.060)
And at that point, when you really see it
Lex Fridman (17:33.900)
and see how little means sometimes the system has
Regina Barzilay (17:38.060)
to have them, you really feel that we need to take a lot
Lex Fridman (17:41.740)
of our brilliance that we have here at MIT
Lex Fridman (17:45.420)
and translate it into something useful.
Lex Fridman (17:48.020)
Yeah, and you still couldn't have a lot of definitions,
Lex Fridman (17:50.540)
but of course, alleviating, suffering, alleviating,
Lex Fridman (17:53.620)
trying to cure cancer is a beautiful mission.
Lex Fridman (17:57.460)
So I of course know theoretically the notion of cancer,
Lex Fridman (18:01.940)
but just reading more and more about it's 1.7 million
Regina Barzilay (18:07.100)
new cancer cases in the United States every year,
Lex Fridman (18:09.860)
600,000 cancer related deaths every year.
Lex Fridman (18:13.460)
So this has a huge impact, United States globally.
Lex Fridman (18:19.340)
When broadly, before we talk about how machine learning,
Lex Fridman (18:24.340)
how MIT can help,
Lex Fridman (18:27.180)
when do you think we as a civilization will cure cancer?
Lex Fridman (18:32.100)
How hard of a problem is it from everything you've learned
Lex Fridman (18:34.980)
from it recently?
Regina Barzilay (18:37.260)
I cannot really assess it.
Lex Fridman (18:39.300)
What I do believe will happen with the advancement
Regina Barzilay (18:42.100)
in machine learning is that a lot of types of cancer
Lex Fridman (18:45.940)
we will be able to predict way early
Lex Fridman (18:48.500)
and more effectively utilize existing treatments.
Lex Fridman (18:53.420)
I think, I hope at least that with all the advancements
Regina Barzilay (18:57.540)
in AI and drug discovery, we would be able
Lex Fridman (19:01.180)
to much faster find relevant molecules.
Lex Fridman (19:04.700)
What I'm not sure about is how long it will take
Lex Fridman (19:08.220)
the medical establishment and regulatory bodies
Regina Barzilay (19:11.940)
to kind of catch up and to implement it.
Lex Fridman (19:14.780)
And I think this is a very big piece of puzzle
Regina Barzilay (19:17.420)
that is currently not addressed.
Lex Fridman (19:20.420)
That's the really interesting question.
Lex Fridman (19:21.780)
So first a small detail that I think the answer is yes,
Lex Fridman (19:25.460)
but is cancer one of the diseases that when detected earlier
Lex Fridman (19:33.700)
that's a significantly improves the outcomes?
Lex Fridman (19:37.820)
So like, cause we will talk about there's the cure
Lex Fridman (19:41.020)
and then there is detection.
Lex Fridman (19:43.020)
And I think where machine learning can really help
Regina Barzilay (19:45.180)
is earlier detection.
Lex Fridman (19:46.660)
So does detection help?
Regina Barzilay (19:48.580)
Detection is crucial.
Lex Fridman (19:49.660)
For instance, the vast majority of pancreatic cancer patients
Regina Barzilay (19:53.940)
are detected at the stage that they are incurable.
Lex Fridman (19:57.300)
That's why they have such a terrible survival rate.
Regina Barzilay (1:00:00.820)
that machine understands you and you can complete the rest
Lex Fridman (1:00:03.940)
that he kind of stopped this research
Lex Fridman (1:00:05.420)
and went into kind of trying to understand
Lex Fridman (1:00:08.660)
what this artificial intelligence can do to our brains.
Lex Fridman (1:00:12.740)
So my point is, you know,
Lex Fridman (1:00:14.380)
how much, it's not how good is the technology,
Regina Barzilay (1:00:19.300)
it's how ready we are to believe
Lex Fridman (1:00:22.620)
that it delivers the goods that we are trying to get.
Regina Barzilay (1:00:25.580)
That's a really beautiful way to put it.
Lex Fridman (1:00:27.200)
I, by the way, I'm not horrified by that possibility,
Lex Fridman (1:00:29.800)
but inspired by it because,
Lex Fridman (1:00:33.140)
I mean, human connection,
Regina Barzilay (1:00:35.920)
whether it's through language or through love,
Lex Fridman (1:00:39.860)
it seems like it's very amenable to machine learning
Lex Fridman (1:00:44.900)
and the rest is just challenges of psychology.
Lex Fridman (1:00:49.340)
Like you said, the secretaries who enjoy spending hours.
Regina Barzilay (1:00:52.460)
I would say I would describe most of our lives
Lex Fridman (1:00:55.020)
as enjoying spending hours with those we love
Regina Barzilay (1:00:58.020)
for very silly reasons.
Lex Fridman (1:01:00.820)
All we're doing is keyword matching as well.
Lex Fridman (1:01:02.780)
So I'm not sure how much intelligence
Lex Fridman (1:01:05.100)
we exhibit to each other with the people we love
Regina Barzilay (1:01:08.140)
that we're close with.
Lex Fridman (1:01:09.820)
So it's a very interesting point
Regina Barzilay (1:01:12.660)
of what it means to pass the Turing test with language.
Lex Fridman (1:01:16.020)
I think you're right.
Regina Barzilay (1:01:16.860)
In terms of conversation,
Lex Fridman (1:01:18.220)
I think machine translation
Lex Fridman (1:01:21.420)
has very clear performance and improvement, right?
Lex Fridman (1:01:24.420)
What it means to have a fulfilling conversation
Regina Barzilay (1:01:28.020)
is very person dependent and context dependent
Lex Fridman (1:01:32.660)
and so on.
Regina Barzilay (1:01:33.580)
That's, yeah, it's very well put.
Lex Fridman (1:01:36.340)
But in your view, what's a benchmark in natural language,
Regina Barzilay (1:01:40.740)
a test that's just out of reach right now,
Lex Fridman (1:01:43.640)
but we might be able to, that's exciting.
Regina Barzilay (1:01:46.020)
Is it in perfecting machine translation
Lex Fridman (1:01:49.100)
or is there other, is it summarization?
Lex Fridman (1:01:51.900)
What's out there just out of reach?
Lex Fridman (1:01:52.740)
I think it goes across specific application.
Regina Barzilay (1:01:55.820)
It's more about the ability to learn from few examples
Lex Fridman (1:01:59.500)
for real, what we call few short learning and all these cases
Regina Barzilay (1:02:03.300)
because the way we publish these papers today,
Lex Fridman (1:02:05.940)
we say, if we have like naively, we get 55,
Lex Fridman (1:02:09.900)
but now we had a few example and we can move to 65.
Lex Fridman (1:02:12.500)
None of these methods
Regina Barzilay (1:02:13.540)
actually are realistically doing anything useful.
Lex Fridman (1:02:15.980)
You cannot use them today.
Lex Fridman (1:02:18.540)
And the ability to be able to generalize and to move
Lex Fridman (1:02:25.460)
or to be autonomous in finding the data
Regina Barzilay (1:02:28.940)
that you need to learn,
Lex Fridman (1:02:31.340)
to be able to perfect new tasks or new language,
Regina Barzilay (1:02:35.300)
this is an area where I think we really need
Lex Fridman (1:02:39.200)
to move forward to and we are not yet there.
Regina Barzilay (1:02:43.020)
Are you at all excited,
Lex Fridman (1:02:45.060)
curious by the possibility
Lex Fridman (1:02:46.540)
of creating human level intelligence?
Lex Fridman (1:02:49.900)
Is this, cause you've been very in your discussion.
Lex Fridman (1:02:52.540)
So if we look at oncology,
Lex Fridman (1:02:54.340)
you're trying to use machine learning to help the world
Regina Barzilay (1:02:58.100)
in terms of alleviating suffering.
Lex Fridman (1:02:59.700)
If you look at natural language processing,
Regina Barzilay (1:03:02.340)
you're focused on the outcomes of improving practical things
Lex Fridman (1:03:05.300)
like machine translation.
Lex Fridman (1:03:06.820)
But human level intelligence is a thing
Lex Fridman (1:03:09.880)
that our civilization has dreamed about creating,
Regina Barzilay (1:03:13.800)
super human level intelligence.
Lex Fridman (1:03:15.740)
Do you think about this?
Lex Fridman (1:03:16.940)
Do you think it's at all within our reach?
Lex Fridman (1:03:20.380)
So as you said yourself, Elie,
Regina Barzilay (1:03:22.660)
talking about how do you perceive
Lex Fridman (1:03:26.140)
our communications with each other,
Regina Barzilay (1:03:28.980)
that we're matching keywords and certain behaviors
Lex Fridman (1:03:31.940)
and so on.
Lex Fridman (1:03:33.020)
So at the end, whenever one assesses,
Lex Fridman (1:03:36.860)
let's say relations with another person,
Regina Barzilay (1:03:38.680)
you have separate kind of measurements and outcomes
Lex Fridman (1:03:41.460)
inside your head that determine
Lex Fridman (1:03:43.620)
what is the status of the relation.
Lex Fridman (1:03:45.860)
So one way, this is this classical level,
Lex Fridman (1:03:48.580)
what is the intelligence?
Lex Fridman (1:03:49.600)
Is it the fact that now we are gonna do the same way
Regina Barzilay (1:03:51.860)
as human is doing,
Lex Fridman (1:03:52.940)
when we don't even understand what the human is doing?
Regina Barzilay (1:03:55.500)
Or we now have an ability to deliver these outcomes,
Lex Fridman (1:03:59.100)
but not in one area, not in NLP,
Regina Barzilay (1:04:01.300)
not just to translate or just to answer questions,
Lex Fridman (1:04:03.940)
but across many, many areas
Regina Barzilay (1:04:05.380)
that we can achieve the functionalities
Lex Fridman (1:04:08.100)
that humans can achieve with their ability to learn
Lex Fridman (1:04:11.060)
and do other things.
Lex Fridman (1:04:12.380)
I think this is, and this we can actually measure
Lex Fridman (1:04:15.500)
how far we are.
Lex Fridman (1:04:17.560)
And that's what makes me excited that we,
Regina Barzilay (1:04:21.580)
in my lifetime, at least so far what we've seen,
Lex Fridman (1:04:23.780)
it's like tremendous progress
Regina Barzilay (1:04:25.840)
across these different functionalities.
Lex Fridman (1:04:28.700)
And I think it will be really exciting
Regina Barzilay (1:04:32.260)
to see where we will be.
Lex Fridman (1:04:35.540)
And again, one way to think about it,
Regina Barzilay (1:04:39.300)
there are machines which are improving their functionality.
Lex Fridman (1:04:41.820)
Another one is to think about us with our brains,
Regina Barzilay (1:04:44.940)
which are imperfect,
Lex Fridman (1:04:46.420)
how they can be accelerated by this technology
Regina Barzilay (1:04:51.420)
as it becomes stronger and stronger.
Lex Fridman (1:04:55.900)
Coming back to another book
Regina Barzilay (1:04:57.260)
that I love, Flowers for Algernon.
Lex Fridman (1:05:01.060)
Have you read this book?
Regina Barzilay (1:05:02.100)
Yes.
Lex Fridman (1:05:02.940)
So there is this point that the patient gets
Regina Barzilay (1:05:05.700)
this miracle cure, which changes his brain.
Lex Fridman (1:05:07.980)
And all of a sudden they see life in a different way
Lex Fridman (1:05:11.020)
and can do certain things better,
Lex Fridman (1:05:13.300)
but certain things much worse.
Lex Fridman (1:05:14.860)
So you can imagine this kind of computer augmented cognition
Lex Fridman (1:05:22.400)
where it can bring you that now in the same way
Regina Barzilay (1:05:24.800)
as the cars enable us to get to places
Lex Fridman (1:05:28.120)
where we've never been before,
Lex Fridman (1:05:30.080)
can we think differently?
Lex Fridman (1:05:31.640)
Can we think faster?
Lex Fridman (1:05:33.600)
And we already see a lot of it happening
Lex Fridman (1:05:36.680)
in how it impacts us,
Lex Fridman (1:05:38.260)
but I think we have a long way to go there.
Lex Fridman (1:05:42.200)
So that's sort of artificial intelligence
Lex Fridman (1:05:45.040)
and technology affecting our,
Lex Fridman (1:05:47.280)
augmenting our intelligence as humans.
Regina Barzilay (1:05:50.440)
Yesterday, a company called Neuralink announced,
Lex Fridman (1:05:55.520)
they did this whole demonstration.
Regina Barzilay (1:05:56.800)
I don't know if you saw it.
Lex Fridman (1:05:57.980)
It's, they demonstrated brain computer,
Regina Barzilay (1:06:01.000)
brain machine interface,
Lex Fridman (1:06:02.680)
where there's like a sewing machine for the brain.
Lex Fridman (1:06:06.360)
Do you, you know, a lot of that is quite out there
Lex Fridman (1:06:11.120)
in terms of things that some people would say
Regina Barzilay (1:06:14.040)
are impossible, but they're dreamers
Lex Fridman (1:06:16.340)
and want to engineer systems like that.
Lex Fridman (1:06:18.080)
Do you see, based on what you just said,
Lex Fridman (1:06:20.360)
a hope for that more direct interaction with the brain?
Regina Barzilay (1:06:25.120)
I think there are different ways.
Lex Fridman (1:06:27.040)
One is a direct interaction with the brain.
Lex Fridman (1:06:29.000)
And again, there are lots of companies
Lex Fridman (1:06:30.900)
that work in this space
Lex Fridman (1:06:32.280)
and I think there will be a lot of developments.
Lex Fridman (1:06:35.080)
But I'm just thinking that many times
Regina Barzilay (1:06:36.600)
we are not aware of our feelings,
Lex Fridman (1:06:39.080)
of motivation, what drives us.
Regina Barzilay (1:06:41.400)
Like, let me give you a trivial example, our attention.
Lex Fridman (1:06:45.520)
There are a lot of studies that demonstrate
Regina Barzilay (1:06:47.260)
that it takes a while to a person to understand
Lex Fridman (1:06:49.200)
that they are not attentive anymore.
Lex Fridman (1:06:51.080)
And we know that there are people
Lex Fridman (1:06:52.160)
who really have strong capacity to hold attention.
Regina Barzilay (1:06:54.520)
There are other end of the spectrum people with ADD
Lex Fridman (1:06:57.080)
and other issues that they have problem
Regina Barzilay (1:06:58.800)
to regulate their attention.
Lex Fridman (1:07:00.760)
Imagine to yourself that you have like a cognitive aid
Regina Barzilay (1:07:03.520)
that just alerts you based on your gaze,
Lex Fridman (1:07:06.280)
that your attention is now not on what you are doing.
Lex Fridman (1:07:09.280)
And instead of writing a paper,
Lex Fridman (1:07:10.560)
you're now dreaming of what you're gonna do in the evening.
Lex Fridman (1:07:12.760)
So even this kind of simple measurement things,
Lex Fridman (1:07:16.360)
how they can change us.
Lex Fridman (1:07:17.840)
And I see it even in simple ways with myself.
Lex Fridman (1:07:22.400)
I have my zone app that I got in MIT gym.
Regina Barzilay (1:07:26.480)
It kind of records, you know, how much did you run
Lex Fridman (1:07:28.800)
and you have some points
Lex Fridman (1:07:29.800)
and you can get some status, whatever.
Lex Fridman (1:07:32.880)
Like, I said, what is this ridiculous thing?
Lex Fridman (1:07:35.840)
Who would ever care about some status in some app?
Lex Fridman (1:07:38.800)
Guess what?
Lex Fridman (1:07:39.640)
So to maintain the status,
Lex Fridman (1:07:41.560)
you have to do set a number of points every month.
Lex Fridman (1:07:44.640)
And not only is that I do it every single month
Lex Fridman (1:07:48.040)
for the last 18 months,
Regina Barzilay (1:07:50.560)
it went to the point that I was injured.
Lex Fridman (1:07:54.160)
And when I could run again,
Regina Barzilay (1:07:58.120)
in two days, I did like some humongous amount of running
Lex Fridman (1:08:02.560)
just to complete the points.
Regina Barzilay (1:08:04.080)
It was like really not safe.
Lex Fridman (1:08:05.920)
It was like, I'm not gonna lose my status
Regina Barzilay (1:08:08.440)
because I want to get there.
Lex Fridman (1:08:10.240)
So you can already see that this direct measurement
Lex Fridman (1:08:13.320)
and the feedback is, you know,
Lex Fridman (1:08:15.160)
we're looking at video games
Lex Fridman (1:08:16.320)
and see why, you know, the addiction aspect of it,
Lex Fridman (1:08:18.720)
but you can imagine that the same idea can be expanded
Regina Barzilay (1:08:21.200)
to many other areas of our life.
Lex Fridman (1:08:23.640)
When we really can get feedback
Lex Fridman (1:08:25.960)
and imagine in your case in relations,
Lex Fridman (1:08:29.880)
when we are doing keyword matching,
Regina Barzilay (1:08:31.240)
imagine that the person who is generating the keywords,
Lex Fridman (1:08:36.120)
that person gets direct feedback
Regina Barzilay (1:08:37.720)
before the whole thing explodes.
Lex Fridman (1:08:39.560)
Is it maybe at this happy point,
Regina Barzilay (1:08:42.000)
we are going in the wrong direction.
Lex Fridman (1:08:44.000)
Maybe it will be really a behavior modifying moment.
Lex Fridman (1:08:48.040)
So yeah, it's a relationship management too.
Lex Fridman (1:08:51.360)
So yeah, that's a fascinating whole area
Regina Barzilay (1:08:54.200)
of psychology actually as well,
Lex Fridman (1:08:56.120)
of seeing how our behavior has changed
Regina Barzilay (1:08:58.240)
with basically all human relations now have
Lex Fridman (1:09:01.840)
other nonhuman entities helping us out.
Lex Fridman (1:09:06.200)
So you teach a large,
Lex Fridman (1:09:09.440)
a huge machine learning course here at MIT.
Regina Barzilay (1:09:14.000)
I can ask you a million questions,
Lex Fridman (1:09:15.360)
but you've seen a lot of students.
Lex Fridman (1:09:17.560)
What ideas do students struggle with the most
Lex Fridman (1:09:20.920)
as they first enter this world of machine learning?
Regina Barzilay (1:09:23.920)
Actually, this year was the first time
Lex Fridman (1:09:26.520)
I started teaching a small machine learning class.
Lex Fridman (1:09:28.480)
And it came as a result of what I saw
Lex Fridman (1:09:31.160)
in my big machine learning class that Tomi Yakel and I built
Regina Barzilay (1:09:34.640)
maybe six years ago.
Lex Fridman (1:09:38.040)
What we've seen that as this area become more
Lex Fridman (1:09:40.360)
and more popular, more and more people at MIT
Lex Fridman (1:09:43.440)
want to take this class.
Lex Fridman (1:09:45.360)
And while we designed it for computer science majors,
Lex Fridman (1:09:48.320)
there were a lot of people who really are interested
Regina Barzilay (1:09:50.760)
to learn it, but unfortunately,
Lex Fridman (1:09:52.600)
their background was not enabling them
Regina Barzilay (1:09:55.720)
to do well in the class.
Lex Fridman (1:09:57.200)
And many of them associated machine learning
Regina Barzilay (1:09:59.360)
with the word struggle and failure,
Lex Fridman (1:10:02.480)
primarily for non majors.
Lex Fridman (1:10:04.640)
And that's why we actually started a new class
Lex Fridman (1:10:06.840)
which we call machine learning from algorithms to modeling,
Regina Barzilay (1:10:10.800)
which emphasizes more the modeling aspects of it
Lex Fridman (1:10:15.000)
and focuses on, it has majors and non majors.
Lex Fridman (1:10:20.000)
So we kind of try to extract the relevant parts
Lex Fridman (1:10:23.480)
and make it more accessible,
Regina Barzilay (1:10:25.560)
because the fact that we're teaching 20 classifiers
Lex Fridman (1:10:27.800)
in standard machine learning class,
Regina Barzilay (1:10:29.240)
it's really a big question to really need it.
Lex Fridman (1:10:32.200)
But it was interesting to see this
Regina Barzilay (1:10:34.520)
from first generation of students,
Lex Fridman (1:10:36.480)
when they came back from their internships
Lex Fridman (1:10:39.080)
and from their jobs,
Lex Fridman (1:10:42.320)
what different and exciting things they can do.
Regina Barzilay (1:10:45.560)
I would never think that you can even apply
Lex Fridman (1:10:47.600)
machine learning to, some of them are like matching,
Regina Barzilay (1:10:50.800)
the relations and other things like variety.
Lex Fridman (1:10:53.480)
Everything is amenable as the machine learning.
Regina Barzilay (1:10:56.080)
That actually brings up an interesting point
Lex Fridman (1:10:58.320)
of computer science in general.
Regina Barzilay (1:11:00.680)
It almost seems, maybe I'm crazy,
Lex Fridman (1:11:03.520)
but it almost seems like everybody needs to learn
Lex Fridman (1:11:06.520)
how to program these days.
Lex Fridman (1:11:08.160)
If you're 20 years old, or if you're starting school,
Regina Barzilay (1:11:11.400)
even if you're an English major,
Lex Fridman (1:11:14.200)
it seems like programming unlocks so much possibility
Regina Barzilay (1:11:20.480)
in this world.
Lex Fridman (1:11:21.880)
So when you interacted with those non majors,
Regina Barzilay (1:11:25.000)
is there skills that they were simply lacking at the time
Lex Fridman (1:11:30.280)
that you wish they had and that they learned
Lex Fridman (1:11:33.000)
in high school and so on?
Lex Fridman (1:11:34.680)
Like how should education change
Lex Fridman (1:11:37.520)
in this computerized world that we live in?
Lex Fridman (1:11:41.320)
I think because I knew that there is a Python component
Regina Barzilay (1:11:44.320)
in the class, their Python skills were okay
Lex Fridman (1:11:47.000)
and the class isn't really heavy on programming.
Regina Barzilay (1:11:49.160)
They primarily kind of add parts to the programs.
Lex Fridman (1:11:52.400)
I think it was more of the mathematical barriers
Lex Fridman (1:11:55.440)
and the class, again, with the design on the majors
Lex Fridman (1:11:58.200)
was using the notation, like big O for complexity
Lex Fridman (1:12:01.200)
and others, people who come from different backgrounds
Lex Fridman (1:12:04.520)
just don't have it in the lexical,
Lex Fridman (1:12:05.800)
so necessarily very challenging notion,
Lex Fridman (1:12:09.120)
but they were just not aware.
Lex Fridman (1:12:12.360)
So I think that kind of linear algebra and probability,
Lex Fridman (1:12:16.240)
the basics, the calculus, multivariate calculus,
Regina Barzilay (1:12:19.120)
things that can help.
Lex Fridman (1:12:20.840)
What advice would you give to students
Regina Barzilay (1:12:23.520)
interested in machine learning,
Lex Fridman (1:12:25.280)
interested, you've talked about detecting,
Regina Barzilay (1:12:29.240)
curing cancer, drug design,
Lex Fridman (1:12:31.360)
if they want to get into that field, what should they do?
Regina Barzilay (1:12:36.320)
Get into it and succeed as researchers
Lex Fridman (1:12:39.040)
and entrepreneurs.
Regina Barzilay (1:12:43.320)
The first good piece of news is that right now
Lex Fridman (1:12:45.240)
there are lots of resources
Regina Barzilay (1:12:47.400)
that are created at different levels
Lex Fridman (1:12:50.160)
and you can find online in your school classes
Regina Barzilay (1:12:54.800)
which are more mathematical, more applied and so on.
Lex Fridman (1:12:57.560)
So you can find a kind of a preacher
Regina Barzilay (1:13:01.320)
which preaches in your own language
Lex Fridman (1:13:02.760)
where you can enter the field
Lex Fridman (1:13:04.520)
and you can make many different types of contribution
Lex Fridman (1:13:06.720)
depending of what is your strengths.
Lex Fridman (1:13:10.760)
And the second point, I think it's really important
Lex Fridman (1:13:13.720)
to find some area which you really care about
Lex Fridman (1:13:18.160)
and it can motivate your learning
Lex Fridman (1:13:20.240)
and it can be for somebody curing cancer
Regina Barzilay (1:13:22.640)
or doing self driving cars or whatever,
Lex Fridman (1:13:25.360)
but to find an area where there is data
Regina Barzilay (1:13:29.680)
where you believe there are strong patterns
Lex Fridman (1:13:31.320)
and we should be doing it and we're still not doing it
Regina Barzilay (1:13:33.600)
or you can do it better
Lex Fridman (1:13:35.280)
and just start there and see where it can bring you.
Lex Fridman (1:13:40.800)
So you've been very successful in many directions in life,
Lex Fridman (1:13:46.480)
but you also mentioned Flowers of Argonon.
Lex Fridman (1:13:51.200)
And I think I've read or listened to you mention somewhere
Lex Fridman (1:13:53.840)
that researchers often get lost
Regina Barzilay (1:13:55.360)
in the details of their work.
Lex Fridman (1:13:56.720)
This is per our original discussion with cancer and so on
Lex Fridman (1:14:00.240)
and don't look at the bigger picture,
Lex Fridman (1:14:02.200)
bigger questions of meaning and so on.
Lex Fridman (1:14:05.320)
So let me ask you the impossible question
Lex Fridman (1:14:08.640)
of what's the meaning of this thing,
Regina Barzilay (1:14:11.560)
of life, of your life, of research.
Lex Fridman (1:14:16.720)
Why do you think we descendant of great apes
Lex Fridman (1:14:21.440)
are here on this spinning ball?
Lex Fridman (1:14:26.800)
You know, I don't think that I have really a global answer.
Regina Barzilay (1:14:30.320)
You know, maybe that's why I didn't go to humanities
Lex Fridman (1:14:33.760)
and I didn't take humanities classes in my undergrad.
Lex Fridman (1:14:39.480)
But the way I'm thinking about it,
Lex Fridman (1:14:43.560)
each one of us inside of them have their own set of,
Regina Barzilay (1:14:48.200)
you know, things that we believe are important.
Lex Fridman (1:14:51.120)
And it just happens that we are busy
Regina Barzilay (1:14:53.360)
with achieving various goals, busy listening to others
Lex Fridman (1:14:56.240)
and to kind of try to conform and to be part of the crowd,
Regina Barzilay (1:15:00.960)
that we don't listen to that part.
Lex Fridman (1:15:04.600)
And, you know, we all should find some time to understand
Lex Fridman (1:15:09.600)
what is our own individual missions.
Lex Fridman (1:15:11.840)
And we may have very different missions
Lex Fridman (1:15:14.080)
and to make sure that while we are running 10,000 things,
Lex Fridman (1:15:18.200)
we are not, you know, missing out
Lex Fridman (1:15:21.920)
and we're putting all the resources to satisfy
Lex Fridman (1:15:26.800)
our own mission.
Lex Fridman (1:15:28.440)
And if I look over my time, when I was younger,
Lex Fridman (1:15:32.400)
most of these missions, you know,
Regina Barzilay (1:15:35.000)
I was primarily driven by the external stimulus,
Lex Fridman (1:15:38.600)
you know, to achieve this or to be that.
Lex Fridman (1:15:41.520)
And now a lot of what I do is driven by really thinking
Lex Fridman (1:15:47.640)
what is important for me to achieve independently
Regina Barzilay (1:15:51.360)
of the external recognition.
Lex Fridman (1:15:55.160)
And, you know, I don't mind to be viewed in certain ways.
Regina Barzilay (1:16:01.400)
The most important thing for me is to be true to myself,
Lex Fridman (1:16:05.760)
to what I think is right.
Lex Fridman (1:16:07.520)
How long did it take?
Lex Fridman (1:16:08.680)
How hard was it to find the you that you have to be true to?
Lex Fridman (1:16:14.160)
So it takes time.
Lex Fridman (1:16:15.520)
And even now, sometimes, you know,
Regina Barzilay (1:16:17.760)
the vanity and the triviality can take, you know.
Lex Fridman (1:16:20.880)
At MIT.
Regina Barzilay (1:16:22.560)
Yeah, it can everywhere, you know,
Lex Fridman (1:16:25.080)
it's just the vanity at MIT is different,
Regina Barzilay (1:16:26.960)
the vanity in different places,
Lex Fridman (1:16:28.160)
but we all have our piece of vanity.
Lex Fridman (1:16:30.920)
But I think actually for me, many times the place
Lex Fridman (1:16:38.720)
to get back to it is, you know, when I'm alone
Lex Fridman (1:16:43.800)
and also when I read.
Lex Fridman (1:16:45.800)
And I think by selecting the right books,
Regina Barzilay (1:16:47.760)
you can get the right questions and learn from what you read.
Lex Fridman (1:16:54.880)
So, but again, it's not perfect.
Regina Barzilay (1:16:58.080)
Like vanity sometimes dominates.
Lex Fridman (1:17:02.040)
Well, that's a beautiful way to end.
Regina Barzilay (1:17:04.800)
Thank you so much for talking today.
Lex Fridman (1:17:06.400)
Thank you.
Regina Barzilay (1:17:07.240)
That was fun.
Lex Fridman (1:17:08.080)
That was fun.
Regina Barzilay (20:03.740)
It's like just few percent over five years.
Lex Fridman (20:07.300)
It's pretty much today the sentence.
Lex Fridman (20:09.820)
But if you can discover this disease early,
Lex Fridman (20:14.500)
there are mechanisms to treat it.
Lex Fridman (20:16.740)
And in fact, I know a number of people who were diagnosed
Lex Fridman (20:20.740)
and saved just because they had food poisoning.
Regina Barzilay (20:23.580)
They had terrible food poisoning.
Lex Fridman (20:25.020)
They went to ER, they got scan.
Regina Barzilay (20:28.540)
There were early signs on the scan
Lex Fridman (20:30.660)
and that would save their lives.
Lex Fridman (20:33.540)
But this wasn't really an accidental case.
Lex Fridman (20:35.820)
So as we become better, we would be able to help
Regina Barzilay (20:41.260)
to many more people that are likely to develop diseases.
Lex Fridman (20:46.540)
And I just want to say that as I got more into this field,
Regina Barzilay (20:51.020)
I realized that cancer is of course terrible disease,
Lex Fridman (20:53.620)
but there are really the whole slew of terrible diseases
Regina Barzilay (20:56.700)
out there like neurodegenerative diseases and others.
Lex Fridman (21:01.660)
So we, of course, a lot of us are fixated on cancer
Regina Barzilay (21:04.580)
because it's so prevalent in our society.
Lex Fridman (21:06.420)
And you see these people where there are a lot of patients
Regina Barzilay (21:08.540)
with neurodegenerative diseases
Lex Fridman (21:10.340)
and the kind of aging diseases
Regina Barzilay (21:12.540)
that we still don't have a good solution for.
Lex Fridman (21:17.100)
And I felt as a computer scientist,
Regina Barzilay (21:22.860)
we kind of decided that it's other people's job
Lex Fridman (21:25.460)
to treat these diseases because it's like traditionally
Regina Barzilay (21:29.340)
people in biology or in chemistry or MDs
Lex Fridman (21:32.420)
are the ones who are thinking about it.
Lex Fridman (21:35.340)
And after kind of start paying attention,
Lex Fridman (21:37.420)
I think that it's really a wrong assumption
Lex Fridman (21:40.340)
and we all need to join the battle.
Lex Fridman (21:42.940)
So how it seems like in cancer specifically
Regina Barzilay (21:46.460)
that there's a lot of ways that machine learning can help.
Lex Fridman (21:49.140)
So what's the role of machine learning
Lex Fridman (21:51.860)
in the diagnosis of cancer?
Lex Fridman (21:55.260)
So for many cancers today, we really don't know
Lex Fridman (21:58.700)
what is your likelihood to get cancer.
Lex Fridman (22:03.460)
And for the vast majority of patients,
Regina Barzilay (22:06.300)
especially on the younger patients,
Lex Fridman (22:07.940)
it really comes as a surprise.
Regina Barzilay (22:09.580)
Like for instance, for breast cancer,
Lex Fridman (22:11.140)
80% of the patients are first in their families,
Regina Barzilay (22:13.860)
it's like me.
Lex Fridman (22:15.380)
And I never saw that I had any increased risk
Regina Barzilay (22:18.460)
because nobody had it in my family.
Lex Fridman (22:20.820)
And for some reason in my head,
Regina Barzilay (22:22.300)
it was kind of inherited disease.
Lex Fridman (22:26.580)
But even if I would pay attention,
Regina Barzilay (22:28.380)
the very simplistic statistical models
Lex Fridman (22:32.420)
that are currently used in clinical practice,
Regina Barzilay (22:34.540)
they really don't give you an answer, so you don't know.
Lex Fridman (22:37.460)
And the same true for pancreatic cancer,
Regina Barzilay (22:40.380)
the same true for non smoking lung cancer and many others.
Lex Fridman (22:45.380)
So what machine learning can do here
Regina Barzilay (22:47.340)
is utilize all this data to tell us early
Lex Fridman (22:51.620)
who is likely to be susceptible
Lex Fridman (22:53.140)
and using all the information that is already there,
Lex Fridman (22:55.980)
be it imaging, be it your other tests,
Lex Fridman (22:59.980)
and eventually liquid biopsies and others,
Lex Fridman (23:04.860)
where the signal itself is not sufficiently strong
Regina Barzilay (23:08.180)
for human eye to do good discrimination
Lex Fridman (23:11.300)
because the signal may be weak,
Lex Fridman (23:12.940)
but by combining many sources,
Lex Fridman (23:15.620)
machine which is trained on large volumes of data
Regina Barzilay (23:18.100)
can really detect it early.
Lex Fridman (23:20.700)
And that's what we've seen with breast cancer
Lex Fridman (23:22.500)
and people are reporting it in other diseases as well.
Lex Fridman (23:25.900)
That really boils down to data, right?
Lex Fridman (23:28.260)
And in the different kinds of sources of data.
Lex Fridman (23:30.980)
And you mentioned regulatory challenges.
Lex Fridman (23:33.740)
So what are the challenges
Lex Fridman (23:35.180)
in gathering large data sets in this space?
Regina Barzilay (23:40.860)
Again, another great question.
Lex Fridman (23:42.660)
So it took me after I decided that I want to work on it
Regina Barzilay (23:45.500)
two years to get access to data.
Lex Fridman (23:48.740)
Any data, like any significant data set?
Regina Barzilay (23:50.580)
Any significant amount, like right now in this country,
Lex Fridman (23:53.580)
there is no publicly available data set
Regina Barzilay (23:57.060)
of modern mammograms that you can just go
Lex Fridman (23:58.820)
on your computer, sign a document and get it.
Regina Barzilay (24:01.860)
It just doesn't exist.
Lex Fridman (24:03.180)
I mean, obviously every hospital has its own collection
Regina Barzilay (24:06.860)
of mammograms.
Lex Fridman (24:07.700)
There are data that came out of clinical trials.
Lex Fridman (24:11.300)
What we're talking about here is a computer scientist
Lex Fridman (24:13.220)
who just wants to run his or her model
Lex Fridman (24:17.140)
and see how it works.
Lex Fridman (24:19.060)
This data, like ImageNet, doesn't exist.
Lex Fridman (24:22.900)
And there is a set which is called like Florida data set
Lex Fridman (24:28.620)
which is a film mammogram from 90s
Regina Barzilay (24:30.860)
which is totally not representative
Lex Fridman (24:32.420)
of the current developments.
Regina Barzilay (24:33.860)
Whatever you're learning on them doesn't scale up.
Lex Fridman (24:35.780)
This is the only resource that is available.
Lex Fridman (24:39.300)
And today there are many agencies
Lex Fridman (24:42.780)
that govern access to data.
Regina Barzilay (24:44.460)
Like the hospital holds your data
Lex Fridman (24:46.300)
and the hospital decides whether they would give it
Regina Barzilay (24:49.260)
to the researcher to work with this data or not.
Lex Fridman (24:52.340)
Individual hospital?
Regina Barzilay (24:54.180)
Yeah.
Lex Fridman (24:55.020)
I mean, the hospital may, you know,
Regina Barzilay (24:57.220)
assuming that you're doing research collaboration,
Lex Fridman (24:59.220)
you can submit, you know,
Regina Barzilay (25:01.980)
there is a proper approval process guided by RB
Lex Fridman (25:05.060)
and if you go through all the processes,
Regina Barzilay (25:07.820)
you can eventually get access to the data.
Lex Fridman (25:10.140)
But if you yourself know our OEI community,
Regina Barzilay (25:13.540)
there are not that many people who actually ever got access
Lex Fridman (25:16.100)
to data because it's very challenging process.
Lex Fridman (25:20.260)
And sorry, just in a quick comment,
Lex Fridman (25:22.780)
MGH or any kind of hospital,
Lex Fridman (25:25.780)
are they scanning the data?
Lex Fridman (25:28.100)
Are they digitally storing it?
Regina Barzilay (25:29.740)
Oh, it is already digitally stored.
Lex Fridman (25:31.580)
You don't need to do any extra processing steps.
Regina Barzilay (25:34.180)
It's already there in the right format is that right now
Lex Fridman (25:38.340)
there are a lot of issues that govern access to the data
Regina Barzilay (25:41.180)
because the hospital is legally responsible for the data.
Lex Fridman (25:46.180)
And, you know, they have a lot to lose
Regina Barzilay (25:51.020)
if they give the data to the wrong person,
Lex Fridman (25:53.140)
but they may not have a lot to gain if they give it
Regina Barzilay (25:56.460)
as a hospital, as a legal entity has given it to you.
Lex Fridman (26:00.580)
And the way, you know, what I would imagine
Regina Barzilay (26:02.740)
happening in the future is the same thing that happens
Lex Fridman (26:05.220)
when you're getting your driving license,
Regina Barzilay (26:06.780)
you can decide whether you want to donate your organs.
Lex Fridman (26:09.820)
You can imagine that whenever a person goes to the hospital,
Regina Barzilay (26:13.100)
they, it should be easy for them to donate their data
Lex Fridman (26:17.540)
for research and it can be different kind of,
Regina Barzilay (26:19.420)
do they only give you a test results or only mammogram
Lex Fridman (26:22.420)
or only imaging data or the whole medical record?
Regina Barzilay (26:27.060)
Because at the end,
Lex Fridman (26:30.540)
we all will benefit from all this insights.
Lex Fridman (26:33.860)
And it's not like you say, I want to keep my data private,
Lex Fridman (26:36.060)
but I would really love to get it from other people
Regina Barzilay (26:38.780)
because other people are thinking the same way.
Lex Fridman (26:40.740)
So if there is a mechanism to do this donation
Lex Fridman (26:45.740)
and the patient has an ability to say
Lex Fridman (26:48.020)
how they want to use their data for research,
Regina Barzilay (26:50.820)
it would be really a game changer.
Lex Fridman (26:54.100)
People, when they think about this problem,
Regina Barzilay (26:56.460)
there's a, it depends on the population,
Lex Fridman (26:58.460)
depends on the demographics,
Lex Fridman (27:00.140)
but there's some privacy concerns generally,
Lex Fridman (27:03.420)
not just medical data, just any kind of data.
Regina Barzilay (27:05.860)
It's what you said, my data, it should belong kind of to me.
Lex Fridman (27:09.620)
I'm worried how it's going to be misused.
Lex Fridman (27:12.540)
How do we alleviate those concerns?
Lex Fridman (27:17.100)
Because that seems like a problem that needs to be,
Regina Barzilay (27:19.460)
that problem of trust, of transparency needs to be solved
Lex Fridman (27:22.980)
before we build large data sets that help detect cancer,
Regina Barzilay (27:27.260)
help save those very people in the future.
Lex Fridman (27:30.180)
So I think there are two things that could be done.
Regina Barzilay (27:31.940)
There is a technical solutions
Lex Fridman (27:34.460)
and there are societal solutions.
Lex Fridman (27:38.220)
So on the technical end,
Lex Fridman (27:41.460)
we today have ability to improve disambiguation.
Regina Barzilay (27:48.140)
Like, for instance, for imaging,
Lex Fridman (27:49.740)
it's, you know, for imaging, you can do it pretty well.
Lex Fridman (27:55.620)
What's disambiguation?
Lex Fridman (27:56.780)
And disambiguation, sorry, disambiguation,
Regina Barzilay (27:58.540)
removing the identification,
Lex Fridman (27:59.860)
removing the names of the people.
Regina Barzilay (28:02.220)
There are other data, like if it is a raw tax,
Lex Fridman (28:04.820)
you cannot really achieve 99.9%,
Lex Fridman (28:08.180)
but there are all these techniques
Lex Fridman (28:10.060)
that actually some of them are developed at MIT,
Lex Fridman (28:12.460)
how you can do learning on the encoded data
Lex Fridman (28:15.460)
where you locally encode the image,
Regina Barzilay (28:17.420)
you train a network which only works on the encoded images
Lex Fridman (28:22.420)
and then you send the outcome back to the hospital
Lex Fridman (28:24.940)
and you can open it up.
Lex Fridman (28:26.580)
So those are the technical solutions.
Regina Barzilay (28:28.020)
There are a lot of people who are working in this space
Lex Fridman (28:30.660)
where the learning happens in the encoded form.
Regina Barzilay (28:33.780)
We are still early,
Lex Fridman (28:36.180)
but this is an interesting research area
Regina Barzilay (28:39.260)
where I think we'll make more progress.
Lex Fridman (28:43.340)
There is a lot of work in natural language processing
Regina Barzilay (28:45.620)
community how to do the identification better.
Lex Fridman (28:50.380)
But even today, there are already a lot of data
Regina Barzilay (28:54.020)
which can be deidentified perfectly,
Lex Fridman (28:55.900)
like your test data, for instance, correct,
Regina Barzilay (28:58.780)
where you can just, you know the name of the patient,
Lex Fridman (29:00.980)
you just want to extract the part with the numbers.
Regina Barzilay (29:04.300)
The big problem here is again,
Lex Fridman (29:08.420)
hospitals don't see much incentive
Regina Barzilay (29:10.420)
to give this data away on one hand
Lex Fridman (29:12.660)
and then there is general concern.
Regina Barzilay (29:14.220)
Now, when I'm talking about societal benefits
Lex Fridman (29:17.700)
and about the education,
Regina Barzilay (29:19.660)
the public needs to understand that I think
Lex Fridman (29:25.700)
that there are situation and I still remember myself
Regina Barzilay (29:29.420)
when I really needed an answer, I had to make a choice.
Lex Fridman (29:33.380)
There was no information to make a choice,
Regina Barzilay (29:35.220)
you're just guessing.
Lex Fridman (29:36.660)
And at that moment you feel that your life is at the stake,
Lex Fridman (29:41.060)
but you just don't have information to make the choice.
Lex Fridman (29:44.820)
And many times when I give talks,
Regina Barzilay (29:48.740)
I get emails from women who say,
Lex Fridman (29:51.300)
you know, I'm in this situation,
Lex Fridman (29:52.820)
can you please run statistic and see what are the outcomes?
Lex Fridman (29:57.100)
We get almost every week a mammogram that comes by mail
Regina Barzilay (30:01.300)
to my office at MIT, I'm serious.
Lex Fridman (30:04.380)
That people ask to run because they need to make
Regina Barzilay (30:07.860)
life changing decisions.
Lex Fridman (30:10.020)
And of course, I'm not planning to open a clinic here,
Lex Fridman (30:12.980)
but we do run and give them the results for their doctors.
Lex Fridman (30:16.660)
But the point that I'm trying to make,
Regina Barzilay (30:20.100)
that we all at some point or our loved ones
Lex Fridman (30:23.780)
will be in the situation where you need information
Regina Barzilay (30:26.620)
to make the best choice.
Lex Fridman (30:28.860)
And if this information is not available,
Regina Barzilay (30:31.860)
you would feel vulnerable and unprotected.
Lex Fridman (30:35.100)
And then the question is, you know, what do I care more?
Lex Fridman (30:37.860)
Because at the end, everything is a trade off, correct?
Lex Fridman (30:40.380)
Yeah, exactly.
Regina Barzilay (30:41.700)
Just out of curiosity, it seems like one possible solution,
Lex Fridman (30:45.580)
I'd like to see what you think of it,
Regina Barzilay (30:49.340)
based on what you just said,
Lex Fridman (30:50.660)
based on wanting to know answers
Regina Barzilay (30:52.500)
for when you're yourself in that situation.
Lex Fridman (30:55.060)
Is it possible for patients to own their data
Lex Fridman (30:58.420)
as opposed to hospitals owning their data?
Lex Fridman (31:01.020)
Of course, theoretically, I guess patients own their data,
Lex Fridman (31:04.100)
but can you walk out there with a USB stick
Lex Fridman (31:07.580)
containing everything or upload it to the cloud?
Regina Barzilay (31:10.620)
Where a company, you know, I remember Microsoft
Lex Fridman (31:14.500)
had a service, like I try, I was really excited about
Lex Fridman (31:17.820)
and Google Health was there.
Lex Fridman (31:19.260)
I tried to give, I was excited about it.
Regina Barzilay (31:21.900)
Basically companies helping you upload your data
Lex Fridman (31:24.780)
to the cloud so that you can move from hospital to hospital
Regina Barzilay (31:27.940)
from doctor to doctor.
Lex Fridman (31:29.260)
Do you see a promise of that kind of possibility?
Regina Barzilay (31:32.700)
I absolutely think this is, you know,
Lex Fridman (31:34.660)
the right way to exchange the data.
Regina Barzilay (31:38.180)
I don't know now who's the biggest player in this field,
Lex Fridman (31:41.700)
but I can clearly see that even for totally selfish
Regina Barzilay (31:45.940)
health reasons, when you are going to a new facility
Lex Fridman (31:49.300)
and many of us are sent to some specialized treatment,
Regina Barzilay (31:52.620)
they don't easily have access to your data.
Lex Fridman (31:55.740)
And today, you know, we might want to send this mammogram,
Regina Barzilay (31:59.420)
need to go to the hospital, find some small office
Lex Fridman (32:01.780)
which gives them the CD and they ship as a CD.
Lex Fridman (32:04.820)
So you can imagine we're looking at kind of decades old
Lex Fridman (32:08.340)
mechanism of data exchange.
Lex Fridman (32:11.340)
So I definitely think this is an area where hopefully
Lex Fridman (32:15.620)
all the right regulatory and technical forces will align
Lex Fridman (32:20.380)
and we will see it actually implemented.
Lex Fridman (32:23.220)
It's sad because unfortunately, and I need to research
Lex Fridman (32:27.500)
why that happened, but I'm pretty sure Google Health
Lex Fridman (32:30.620)
and Microsoft Health Vault or whatever it's called
Regina Barzilay (32:32.940)
both closed down, which means that there was
Lex Fridman (32:36.100)
either regulatory pressure or there's not a business case
Regina Barzilay (32:39.100)
or there's challenges from hospitals,
Lex Fridman (32:41.820)
which is very disappointing.
Lex Fridman (32:43.260)
So when you say you don't know what the biggest players are,
Lex Fridman (32:46.500)
the two biggest that I was aware of closed their doors.
Lex Fridman (32:50.540)
So I'm hoping, I'd love to see why
Lex Fridman (32:53.140)
and I'd love to see who else can come up.
Regina Barzilay (32:54.780)
It seems like one of those Elon Musk style problems
Lex Fridman (32:59.620)
that are obvious needs to be solved
Lex Fridman (33:01.300)
and somebody needs to step up and actually do
Lex Fridman (33:02.980)
this large scale data collection.
Lex Fridman (33:07.540)
So I know there is an initiative in Massachusetts,
Lex Fridman (33:09.620)
I think, which you led by the governor
Regina Barzilay (33:11.740)
to try to create this kind of health exchange system
Lex Fridman (33:15.460)
where at least to help people who kind of when you show up
Regina Barzilay (33:17.860)
in emergency room and there is no information
Lex Fridman (33:20.220)
about what are your allergies and other things.
Lex Fridman (33:23.540)
So I don't know how far it will go.
Lex Fridman (33:26.140)
But another thing that you said
Lex Fridman (33:28.180)
and I find it very interesting is actually
Lex Fridman (33:30.780)
who are the successful players in this space
Lex Fridman (33:33.780)
and the whole implementation, how does it go?
Lex Fridman (33:37.260)
To me, it is from the anthropological perspective,
Regina Barzilay (33:40.300)
it's more fascinating that AI that today goes in healthcare,
Lex Fridman (33:44.660)
we've seen so many attempts and so very little successes.
Lex Fridman (33:50.380)
And it's interesting to understand that I've by no means
Lex Fridman (33:54.220)
have knowledge to assess it,
Lex Fridman (33:56.700)
why we are in the position where we are.
Lex Fridman (33:59.620)
Yeah, it's interesting because data is really fuel
Regina Barzilay (34:02.940)
for a lot of successful applications.
Lex Fridman (34:04.980)
And when that data acquires regulatory approval,
Regina Barzilay (34:08.500)
like the FDA or any kind of approval,
Lex Fridman (34:12.940)
it seems that the computer scientists
Regina Barzilay (34:15.740)
are not quite there yet in being able
Lex Fridman (34:17.460)
to play the regulatory game,
Regina Barzilay (34:18.900)
understanding the fundamentals of it.
Lex Fridman (34:21.220)
I think that in many cases when even people do have data,
Regina Barzilay (34:26.500)
we still don't know what exactly do you need to demonstrate
Lex Fridman (34:31.300)
to change the standard of care.
Regina Barzilay (34:35.500)
Like let me give you an example
Lex Fridman (34:37.180)
related to my breast cancer research.
Lex Fridman (34:41.100)
So in traditional breast cancer risk assessment,
Lex Fridman (34:45.500)
there is something called density,
Regina Barzilay (34:47.140)
which determines the likelihood of a woman to get cancer.
Lex Fridman (34:50.500)
And this pretty much says,
Lex Fridman (34:51.700)
how much white do you see on the mammogram?
Lex Fridman (34:54.220)
The whiter it is, the more likely the tissue is dense.
Lex Fridman (34:58.980)
And the idea behind density, it's not a bad idea.
Lex Fridman (35:03.660)
In 1967, a radiologist called Wolf decided to look back
Regina Barzilay (35:08.100)
at women who were diagnosed
Lex Fridman (35:09.780)
and see what is special in their images.
Lex Fridman (35:12.420)
Can we look back and say that they're likely to develop?
Lex Fridman (35:14.700)
So he come up with some patterns.
Lex Fridman (35:16.180)
And it was the best that his human eye can identify.
Lex Fridman (35:20.660)
Then it was kind of formalized
Lex Fridman (35:22.060)
and coded into four categories.
Lex Fridman (35:24.220)
And that's what we are using today.
Lex Fridman (35:26.940)
And today this density assessment
Lex Fridman (35:31.020)
is actually a federal law from 2019,
Regina Barzilay (35:34.620)
approved by President Trump
Lex Fridman (35:36.180)
and for the previous FDA commissioner,
Regina Barzilay (35:40.100)
where women are supposed to be advised by their providers
Lex Fridman (35:43.620)
if they have high density,
Regina Barzilay (35:45.100)
putting them into higher risk category.
Lex Fridman (35:47.260)
And in some states,
Regina Barzilay (35:49.460)
you can actually get supplementary screening
Lex Fridman (35:51.260)
paid by your insurance because you're in this category.
Lex Fridman (35:53.700)
Now you can say, how much science do we have behind it?
Lex Fridman (35:56.780)
Whatever, biological science or epidemiological evidence.
Lex Fridman (36:00.820)
So it turns out that between 40 and 50% of women
Lex Fridman (36:05.140)
have dense breasts.
Lex Fridman (36:06.660)
So about 40% of patients are coming out of their screening
Lex Fridman (36:11.140)
and somebody tells them, you are in high risk.
Regina Barzilay (36:15.020)
Now, what exactly does it mean
Lex Fridman (36:16.860)
if you as half of the population in high risk?
Regina Barzilay (36:19.620)
It's from saying, maybe I'm not,
Lex Fridman (36:22.060)
or what do I really need to do with it?
Regina Barzilay (36:23.700)
Because the system doesn't provide me
Lex Fridman (36:27.220)
a lot of the solutions
Regina Barzilay (36:28.340)
because there are so many people like me,
Lex Fridman (36:30.140)
we cannot really provide very expensive solutions for them.
Lex Fridman (36:34.620)
And the reason this whole density became this big deal,
Lex Fridman (36:38.740)
it's actually advocated by the patients
Regina Barzilay (36:40.820)
who felt very unprotected
Lex Fridman (36:42.500)
because many women went and did the mammograms
Regina Barzilay (36:44.900)
which were normal.
Lex Fridman (36:46.260)
And then it turns out that they already had cancer,
Regina Barzilay (36:49.460)
quite developed cancer.
Lex Fridman (36:50.580)
So they didn't have a way to know who is really at risk
Lex Fridman (36:54.420)
and what is the likelihood that when the doctor tells you,
Lex Fridman (36:56.300)
you're okay, you are not okay.
Lex Fridman (36:58.060)
So at the time, and it was 15 years ago,
Lex Fridman (37:02.140)
this maybe was the best piece of science that we had.
Lex Fridman (37:06.820)
And it took quite 15, 16 years to make it federal law.
Lex Fridman (37:12.180)
But now this is a standard.
Regina Barzilay (37:15.660)
Now with a deep learning model,
Lex Fridman (37:17.620)
we can so much more accurately predict
Regina Barzilay (37:19.660)
who is gonna develop breast cancer
Lex Fridman (37:21.580)
just because you're trained on a logical thing.
Lex Fridman (37:23.700)
And instead of describing how much white
Lex Fridman (37:26.060)
and what kind of white machine
Regina Barzilay (37:27.380)
can systematically identify the patterns,
Lex Fridman (37:30.140)
which was the original idea behind the thought
Regina Barzilay (37:32.780)
of the cardiologist,
Lex Fridman (37:33.700)
machines can do it much more systematically
Lex Fridman (37:35.740)
and predict the risk when you're training the machine
Lex Fridman (37:38.260)
to look at the image and to say the risk in one to five years.
Regina Barzilay (37:42.140)
Now you can ask me how long it will take
Lex Fridman (37:45.060)
to substitute this density,
Regina Barzilay (37:46.460)
which is broadly used across the country
Lex Fridman (37:48.620)
and really is not helping to bring this new models.
Lex Fridman (37:54.380)
And I would say it's not a matter of the algorithm.
Lex Fridman (37:56.700)
Algorithms use already orders of magnitude better
Regina Barzilay (37:58.780)
than what is currently in practice.
Lex Fridman (38:00.460)
I think it's really the question,
Lex Fridman (38:02.500)
who do you need to convince?
Lex Fridman (38:04.380)
How many hospitals do you need to run the experiment?
Regina Barzilay (38:07.460)
What, you know, all this mechanism of adoption
Lex Fridman (38:11.500)
and how do you explain to patients
Lex Fridman (38:15.180)
and to women across the country
Lex Fridman (38:17.580)
that this is really a better measure?
Lex Fridman (38:20.460)
And again, I don't think it's an AI question.
Lex Fridman (38:22.740)
We can work more and make the algorithm even better,
Lex Fridman (38:25.940)
but I don't think that this is the current, you know,
Lex Fridman (38:29.300)
the barrier, the barrier is really this other piece
Regina Barzilay (38:32.060)
that for some reason is not really explored.
Lex Fridman (38:35.260)
It's like anthropological piece.
Lex Fridman (38:36.860)
And coming back to your question about books,
Lex Fridman (38:39.860)
there is a book that I'm reading.
Regina Barzilay (38:42.980)
It's called American Sickness by Elizabeth Rosenthal.
Lex Fridman (38:48.260)
And I got this book from my clinical collaborator,
Regina Barzilay (38:51.580)
Dr. Connie Lehman.
Lex Fridman (38:53.100)
And I said, I know everything that I need to know
Regina Barzilay (38:54.820)
about American health system,
Lex Fridman (38:56.020)
but you know, every page doesn't fail to surprise me.
Lex Fridman (38:59.220)
And I think there is a lot of interesting
Lex Fridman (39:03.140)
and really deep lessons for people like us
Regina Barzilay (39:06.860)
from computer science who are coming into this field
Lex Fridman (39:09.660)
to really understand how complex is the system of incentives
Regina Barzilay (39:13.660)
in the system to understand how you really need to play
Lex Fridman (39:17.660)
to drive adoption.
Regina Barzilay (39:19.740)
You just said it's complex,
Lex Fridman (39:21.180)
but if we're trying to simplify it,
Regina Barzilay (39:23.980)
who do you think most likely would be successful
Lex Fridman (39:27.380)
if we push on this group of people?
Lex Fridman (39:29.540)
Is it the doctors?
Lex Fridman (39:30.780)
Is it the hospitals?
Lex Fridman (39:31.820)
Is it the governments or policymakers?
Lex Fridman (39:34.300)
Is it the individual patients, consumers?
Lex Fridman (39:38.860)
Who needs to be inspired to most likely lead to adoption?
Lex Fridman (39:45.180)
Or is there no simple answer?
Regina Barzilay (39:47.100)
There's no simple answer,
Lex Fridman (39:48.260)
but I think there is a lot of good people in medical system
Regina Barzilay (39:51.980)
who do want to make a change.
Lex Fridman (39:56.460)
And I think a lot of power will come from us as consumers
Regina Barzilay (40:01.540)
because we all are consumers or future consumers
Lex Fridman (40:04.260)
of healthcare services.
Lex Fridman (40:06.500)
And I think we can do so much more
Lex Fridman (40:12.060)
in explaining the potential and not in the hype terms
Lex Fridman (40:15.500)
and not saying that we now killed all Alzheimer
Lex Fridman (40:17.900)
and I'm really sick of reading this kind of articles
Regina Barzilay (40:20.500)
which make these claims,
Lex Fridman (40:22.100)
but really to show with some examples
Lex Fridman (40:24.780)
what this implementation does and how it changes the care.
Lex Fridman (40:29.060)
Because I can't imagine,
Regina Barzilay (40:30.020)
it doesn't matter what kind of politician it is,
Lex Fridman (40:33.220)
we all are susceptible to these diseases.
Regina Barzilay (40:35.220)
There is no one who is free.
Lex Fridman (40:37.740)
And eventually, we all are humans
Lex Fridman (40:41.060)
and we're looking for a way to alleviate the suffering.
Lex Fridman (40:44.860)
And this is one possible way
Regina Barzilay (40:47.260)
where we currently are under utilizing,
Lex Fridman (40:49.300)
which I think can help.
Lex Fridman (40:51.860)
So it sounds like the biggest problems are outside of AI
Lex Fridman (40:55.100)
in terms of the biggest impact at this point.
Lex Fridman (40:57.980)
But are there any open problems
Lex Fridman (41:00.420)
in the application of ML to oncology in general?
Lex Fridman (41:03.780)
So improving the detection or any other creative methods,
Lex Fridman (41:07.540)
whether it's on the detection segmentations
Regina Barzilay (41:09.620)
or the vision perception side
Lex Fridman (41:11.780)
or some other clever of inference?
Regina Barzilay (41:16.260)
Yeah, what in general in your view are the open problems
Lex Fridman (41:19.620)
in this space?
Regina Barzilay (41:20.460)
Yeah, I just want to mention that beside detection,
Lex Fridman (41:22.460)
not the area where I am kind of quite active
Lex Fridman (41:24.820)
and I think it's really an increasingly important area
Lex Fridman (41:28.580)
in healthcare is drug design.
Regina Barzilay (41:32.260)
Absolutely.
Lex Fridman (41:33.100)
Because it's fine if you detect something early,
Lex Fridman (41:36.900)
but you still need to get drugs
Lex Fridman (41:41.100)
and new drugs for these conditions.
Lex Fridman (41:43.860)
And today, all of the drug design,
Lex Fridman (41:46.740)
ML is non existent there.
Regina Barzilay (41:48.300)
We don't have any drug that was developed by the ML model
Lex Fridman (41:52.980)
or even not developed,
Lex Fridman (41:54.900)
but at least even knew that ML model
Lex Fridman (41:57.060)
plays some significant role.
Regina Barzilay (41:59.260)
I think this area with all the new ability
Lex Fridman (42:03.300)
to generate molecules with desired properties
Regina Barzilay (42:05.780)
to do in silica screening is really a big open area.
Lex Fridman (42:11.460)
To be totally honest with you,
Regina Barzilay (42:12.740)
when we are doing diagnostics and imaging,
Lex Fridman (42:14.900)
primarily taking the ideas that were developed
Regina Barzilay (42:17.260)
for other areas and you applying them with some adaptation,
Lex Fridman (42:20.460)
the area of drug design is really technically interesting
Lex Fridman (42:26.820)
and exciting area.
Lex Fridman (42:27.980)
You need to work a lot with graphs
Lex Fridman (42:30.380)
and capture various 3D properties.
Lex Fridman (42:34.580)
There are lots and lots of opportunities
Regina Barzilay (42:37.420)
to be technically creative.
Lex Fridman (42:39.820)
And I think there are a lot of open questions in this area.
Regina Barzilay (42:46.820)
We're already getting a lot of successes
Lex Fridman (42:48.820)
even with kind of the first generation of these models,
Lex Fridman (42:52.700)
but there is much more new creative things that you can do.
Lex Fridman (42:56.500)
And what's very nice to see is that actually
Regina Barzilay (42:59.260)
the more powerful, the more interesting models
Lex Fridman (43:04.180)
actually do do better.
Lex Fridman (43:05.460)
So there is a place to innovate in machine learning
Lex Fridman (43:11.300)
in this area.
Lex Fridman (43:13.900)
And some of these techniques are really unique to,
Lex Fridman (43:16.820)
let's say, to graph generation and other things.
Regina Barzilay (43:19.620)
So...
Lex Fridman (43:20.820)
What, just to interrupt really quick, I'm sorry,
Regina Barzilay (43:23.980)
graph generation or graphs, drug discovery in general,
Lex Fridman (43:30.620)
how do you discover a drug?
Lex Fridman (43:31.940)
Is this chemistry?
Lex Fridman (43:33.340)
Is this trying to predict different chemical reactions?
Regina Barzilay (43:37.500)
Or is it some kind of...
Lex Fridman (43:39.660)
What do graphs even represent in this space?
Regina Barzilay (43:42.100)
Oh, sorry, sorry.
Lex Fridman (43:43.980)
And what's a drug?
Regina Barzilay (43:45.340)
Okay, so let's say you're thinking
Lex Fridman (43:47.140)
there are many different types of drugs,
Lex Fridman (43:48.540)
but let's say you're gonna talk about small molecules
Lex Fridman (43:50.580)
because I think today the majority of drugs
Regina Barzilay (43:52.860)
are small molecules.
Lex Fridman (43:53.700)
So small molecule is a graph.
Regina Barzilay (43:55.020)
The molecule is just where the node in the graph
Lex Fridman (43:59.180)
is an atom and then you have the bonds.
Lex Fridman (44:01.500)
So it's really a graph representation.
Lex Fridman (44:03.220)
If you look at it in 2D, correct,
Regina Barzilay (44:05.540)
you can do it 3D, but let's say,
Lex Fridman (44:07.460)
let's keep it simple and stick in 2D.
Lex Fridman (44:11.500)
So pretty much my understanding today,
Lex Fridman (44:14.740)
how it is done at scale in the companies,
Regina Barzilay (44:18.620)
without machine learning,
Lex Fridman (44:20.220)
you have high throughput screening.
Lex Fridman (44:22.100)
So you know that you are interested
Lex Fridman (44:23.740)
to get certain biological activity of the compound.
Lex Fridman (44:26.540)
So you scan a lot of compounds,
Lex Fridman (44:28.860)
like maybe hundreds of thousands,
Regina Barzilay (44:30.700)
some really big number of compounds.
Lex Fridman (44:32.980)
You identify some compounds which have the right activity
Lex Fridman (44:36.060)
and then at this point, the chemists come
Lex Fridman (44:39.220)
and they're trying to now to optimize
Regina Barzilay (44:43.220)
this original heat to different properties
Lex Fridman (44:45.340)
that you want it to be maybe soluble,
Regina Barzilay (44:47.180)
you want it to decrease toxicity,
Lex Fridman (44:49.060)
you want it to decrease the side effects.
Regina Barzilay (44:51.620)
Are those, sorry again to interrupt,
Lex Fridman (44:54.020)
can that be done in simulation
Regina Barzilay (44:55.500)
or just by looking at the molecules
Lex Fridman (44:57.700)
or do you need to actually run reactions
Lex Fridman (44:59.820)
in real labs with lab coats and stuff?
Lex Fridman (45:02.460)
So when you do high throughput screening,
Regina Barzilay (45:04.020)
you really do screening.
Lex Fridman (45:06.100)
It's in the lab.
Regina Barzilay (45:07.020)
It's really the lab screening.
Lex Fridman (45:09.140)
You screen the molecules, correct?
Regina Barzilay (45:10.980)
I don't know what screening is.
Lex Fridman (45:12.580)
The screening is just check them for certain property.
Regina Barzilay (45:15.060)
Like in the physical space, in the physical world,
Lex Fridman (45:17.260)
like actually there's a machine probably
Regina Barzilay (45:18.740)
that's actually running the reaction.
Lex Fridman (45:21.420)
Actually running the reactions, yeah.
Lex Fridman (45:22.900)
So there is a process where you can run
Lex Fridman (45:25.420)
and that's why it's called high throughput
Regina Barzilay (45:26.660)
that it become cheaper and faster
Lex Fridman (45:29.580)
to do it on very big number of molecules.
Regina Barzilay (45:33.820)
You run the screening,
Lex Fridman (45:35.820)
you identify potential good starts
Lex Fridman (45:40.300)
and then when the chemists come in
Lex Fridman (45:42.340)
who have done it many times
Lex Fridman (45:44.060)
and then they can try to look at it and say,
Lex Fridman (45:46.180)
how can you change the molecule
Regina Barzilay (45:48.260)
to get the desired profile
Lex Fridman (45:51.780)
in terms of all other properties?
Lex Fridman (45:53.460)
So maybe how do I make it more bioactive and so on?
Lex Fridman (45:56.500)
And there the creativity of the chemists
Regina Barzilay (45:59.460)
really is the one that determines the success
Lex Fridman (46:03.980)
of this design because again,
Regina Barzilay (46:07.460)
they have a lot of domain knowledge
Lex Fridman (46:09.300)
of what works, how do you decrease the CCD and so on
Lex Fridman (46:12.900)
and that's what they do.
Lex Fridman (46:15.020)
So all the drugs that are currently
Regina Barzilay (46:17.860)
in the FDA approved drugs
Lex Fridman (46:20.220)
or even drugs that are in clinical trials,
Regina Barzilay (46:22.140)
they are designed using these domain experts
Lex Fridman (46:27.100)
which goes through this combinatorial space
Regina Barzilay (46:30.060)
of molecules or graphs or whatever
Lex Fridman (46:31.940)
and find the right one or adjust it to be the right ones.
Regina Barzilay (46:35.140)
It sounds like the breast density heuristic
Lex Fridman (46:38.060)
from 67 to the same echoes.
Regina Barzilay (46:40.460)
It's not necessarily that.
Lex Fridman (46:41.820)
It's really driven by deep understanding.
Regina Barzilay (46:45.380)
It's not like they just observe it.
Lex Fridman (46:46.820)
I mean, they do deeply understand chemistry
Lex Fridman (46:48.540)
and they do understand how different groups
Lex Fridman (46:50.460)
and how does it changes the properties.
Lex Fridman (46:53.140)
So there is a lot of science that gets into it
Lex Fridman (46:56.660)
and a lot of kind of simulation,
Lex Fridman (46:58.740)
how do you want it to behave?
Lex Fridman (47:01.900)
It's very, very complex.
Lex Fridman (47:03.900)
So they're quite effective at this design, obviously.
Lex Fridman (47:06.140)
Now effective, yeah, we have drugs.
Regina Barzilay (47:08.420)
Like depending on how do you measure effective,
Lex Fridman (47:10.780)
if you measure it in terms of cost, it's prohibitive.
Regina Barzilay (47:13.940)
If you measure it in terms of times,
Lex Fridman (47:15.820)
we have lots of diseases for which we don't have any drugs
Lex Fridman (47:18.420)
and we don't even know how to approach
Lex Fridman (47:20.060)
and don't need to mention few drugs
Regina Barzilay (47:23.460)
or neurodegenerative disease drugs that fail.
Lex Fridman (47:27.140)
So there are lots of trials that fail in later stages,
Regina Barzilay (47:32.180)
which is really catastrophic from the financial perspective.
Lex Fridman (47:35.180)
So is it the effective, the most effective mechanism?
Regina Barzilay (47:39.540)
Absolutely no, but this is the only one that currently works.
Lex Fridman (47:44.300)
And I was closely interacting
Regina Barzilay (47:47.900)
with people in pharmaceutical industry.
Lex Fridman (47:49.260)
I was really fascinated on how sharp
Lex Fridman (47:51.340)
and what a deep understanding of the domain do they have.
Lex Fridman (47:55.260)
It's not observation driven.
Regina Barzilay (47:57.020)
There is really a lot of science behind what they do.
Lex Fridman (48:00.220)
But if you ask me, can machine learning change it,
Regina Barzilay (48:02.300)
I firmly believe yes,
Lex Fridman (48:05.300)
because even the most experienced chemists
Regina Barzilay (48:07.860)
cannot hold in their memory and understanding
Lex Fridman (48:11.100)
everything that you can learn
Regina Barzilay (48:12.500)
from millions of molecules and reactions.
Lex Fridman (48:17.220)
And the space of graphs is a totally new space.
Regina Barzilay (48:19.900)
I mean, it's a really interesting space
Lex Fridman (48:22.060)
for machine learning to explore, graph generation.
Regina Barzilay (48:23.980)
Yeah, so there are a lot of things that you can do here.
Lex Fridman (48:26.260)
So we do a lot of work.
Lex Fridman (48:28.740)
So the first tool that we started with
Lex Fridman (48:31.620)
was the tool that can predict properties of the molecules.
Lex Fridman (48:36.300)
So you can just give the molecule and the property.
Lex Fridman (48:39.420)
It can be by activity property,
Regina Barzilay (48:41.340)
or it can be some other property.
Lex Fridman (48:44.300)
And you train the molecules
Lex Fridman (48:46.460)
and you can now take a new molecule
Lex Fridman (48:50.020)
and predict this property.
Regina Barzilay (48:52.180)
Now, when people started working in this area,
Lex Fridman (48:54.860)
it is something very simple.
Regina Barzilay (48:55.980)
They do kind of existing fingerprints,
Lex Fridman (48:58.580)
which is kind of handcrafted features of the molecule.
Regina Barzilay (49:00.740)
When you break the graph to substructures
Lex Fridman (49:02.980)
and then you run it in a feed forward neural network.
Lex Fridman (49:05.980)
And what was interesting to see that clearly,
Lex Fridman (49:08.500)
this was not the most effective way to proceed.
Lex Fridman (49:11.020)
And you need to have much more complex models
Lex Fridman (49:14.060)
that can induce a representation,
Regina Barzilay (49:16.300)
which can translate this graph into the embeddings
Lex Fridman (49:19.220)
and do these predictions.
Lex Fridman (49:21.300)
So this is one direction.
Lex Fridman (49:23.220)
Then another direction, which is kind of related
Regina Barzilay (49:25.260)
is not only to stop by looking at the embedding itself,
Lex Fridman (49:29.180)
but actually modify it to produce better molecules.
Lex Fridman (49:32.780)
So you can think about it as machine translation
Lex Fridman (49:36.020)
that you can start with a molecule
Lex Fridman (49:38.140)
and then there is an improved version of molecule.
Lex Fridman (49:40.580)
And you can again, with encoder translate it
Regina Barzilay (49:42.860)
into the hidden space and then learn how to modify it
Lex Fridman (49:45.380)
to improve the in some ways version of the molecules.
Lex Fridman (49:49.340)
So that's, it's kind of really exciting.
Lex Fridman (49:52.620)
We already have seen that the property prediction
Regina Barzilay (49:54.740)
works pretty well.
Lex Fridman (49:56.140)
And now we are generating molecules
Lex Fridman (49:59.780)
and there is actually labs
Lex Fridman (50:01.820)
which are manufacturing this molecule.
Lex Fridman (50:04.180)
So we'll see where it will get us.
Lex Fridman (50:06.340)
Okay, that's really exciting.
Regina Barzilay (50:07.780)
There's a lot of promise.
Lex Fridman (50:08.860)
Speaking of machine translation and embeddings,
Regina Barzilay (50:11.820)
I think you have done a lot of really great research
Lex Fridman (50:15.580)
in NLP, natural language processing.
Lex Fridman (50:19.260)
Can you tell me your journey through NLP?
Lex Fridman (50:21.540)
What ideas, problems, approaches were you working on?
Regina Barzilay (50:25.100)
Were you fascinated with, did you explore
Lex Fridman (50:28.180)
before this magic of deep learning reemerged and after?
Lex Fridman (50:34.020)
So when I started my work in NLP, it was in 97.
Lex Fridman (50:38.180)
This was very interesting time.
Regina Barzilay (50:39.460)
It was exactly the time that I came to ACL.
Lex Fridman (50:43.500)
And at the time I could barely understand English,
Lex Fridman (50:46.140)
but it was exactly like the transition point
Lex Fridman (50:48.500)
because half of the papers were really rule based approaches
Regina Barzilay (50:53.500)
where people took more kind of heavy linguistic approaches
Lex Fridman (50:56.180)
for small domains and try to build up from there.
Lex Fridman (51:00.060)
And then there were the first generation of papers
Lex Fridman (51:02.220)
which were corpus based papers.
Lex Fridman (51:04.500)
And they were very simple in our terms
Lex Fridman (51:06.420)
when you collect some statistics
Lex Fridman (51:07.900)
and do prediction based on them.
Lex Fridman (51:10.020)
And I found it really fascinating that one community
Regina Barzilay (51:13.100)
can think so very differently about the problem.
Lex Fridman (51:19.220)
And I remember my first paper that I wrote,
Regina Barzilay (51:22.820)
it didn't have a single formula.
Lex Fridman (51:24.460)
It didn't have evaluation.
Regina Barzilay (51:25.740)
It just had examples of outputs.
Lex Fridman (51:28.340)
And this was a standard of the field at the time.
Regina Barzilay (51:32.020)
In some ways, I mean, people maybe just started emphasizing
Lex Fridman (51:35.860)
the empirical evaluation, but for many applications
Regina Barzilay (51:38.940)
like summarization, you just show some examples of outputs.
Lex Fridman (51:42.780)
And then increasingly you can see that how
Regina Barzilay (51:45.460)
the statistical approaches dominated the field
Lex Fridman (51:48.300)
and we've seen increased performance
Regina Barzilay (51:52.100)
across many basic tasks.
Lex Fridman (51:56.020)
The sad part of the story maybe that if you look again
Regina Barzilay (52:00.420)
through this journey, we see that the role of linguistics
Lex Fridman (52:05.100)
in some ways greatly diminishes.
Lex Fridman (52:07.460)
And I think that you really need to look
Lex Fridman (52:11.580)
through the whole proceeding to find one or two papers
Regina Barzilay (52:14.540)
which make some interesting linguistic references.
Lex Fridman (52:17.260)
It's really big.
Regina Barzilay (52:18.100)
Today, yeah.
Lex Fridman (52:18.920)
Today, today.
Regina Barzilay (52:19.760)
This was definitely one of the.
Lex Fridman (52:20.600)
Things like syntactic trees, just even basically
Regina Barzilay (52:23.140)
against our conversation about human understanding
Lex Fridman (52:26.180)
of language, which I guess what linguistics would be
Regina Barzilay (52:30.300)
structured, hierarchical representing language
Lex Fridman (52:34.300)
in a way that's human explainable, understandable
Regina Barzilay (52:37.140)
is missing today.
Lex Fridman (52:39.500)
I don't know if it is, what is explainable
Lex Fridman (52:42.380)
and understandable.
Lex Fridman (52:43.620)
In the end, we perform functions and it's okay
Regina Barzilay (52:47.360)
to have machine which performs a function.
Lex Fridman (52:50.140)
Like when you're thinking about your calculator, correct?
Regina Barzilay (52:53.200)
Your calculator can do calculation very different
Lex Fridman (52:56.100)
from you would do the calculation,
Lex Fridman (52:57.620)
but it's very effective in it.
Lex Fridman (52:58.860)
And this is fine if we can achieve certain tasks
Regina Barzilay (53:02.560)
with high accuracy, doesn't necessarily mean
Lex Fridman (53:05.760)
that it has to understand it the same way as we understand.
Regina Barzilay (53:09.300)
In some ways, it's even naive to request
Lex Fridman (53:11.260)
because you have so many other sources of information
Regina Barzilay (53:14.940)
that are absent when you are training your system.
Lex Fridman (53:17.900)
So it's okay.
Lex Fridman (53:19.220)
Is it delivered?
Lex Fridman (53:20.060)
And I would tell you one application
Regina Barzilay (53:21.500)
that is really fascinating.
Lex Fridman (53:22.780)
In 97, when it came to ACL, there were some papers
Regina Barzilay (53:25.060)
on machine translation.
Lex Fridman (53:25.900)
They were like primitive.
Regina Barzilay (53:27.440)
Like people were trying really, really simple.
Lex Fridman (53:31.060)
And the feeling, my feeling was that, you know,
Regina Barzilay (53:34.260)
to make real machine translation system,
Lex Fridman (53:36.260)
it's like to fly at the moon and build a house there
Lex Fridman (53:39.580)
and the garden and live happily ever after.
Lex Fridman (53:41.580)
I mean, it's like impossible.
Regina Barzilay (53:42.600)
I never could imagine that within, you know, 10 years,
Lex Fridman (53:46.740)
we would already see the system working.
Lex Fridman (53:48.540)
And now, you know, nobody is even surprised
Lex Fridman (53:51.420)
to utilize the system on daily basis.
Lex Fridman (53:54.420)
So this was like a huge, huge progress,
Lex Fridman (53:56.220)
saying that people for very long time
Regina Barzilay (53:57.860)
tried to solve using other mechanisms.
Lex Fridman (54:00.820)
And they were unable to solve it.
Regina Barzilay (54:03.220)
That's why coming back to your question about biology,
Lex Fridman (54:06.140)
that, you know, in linguistics, people try to go this way
Lex Fridman (54:10.800)
and try to write the syntactic trees
Lex Fridman (54:13.500)
and try to abstract it and to find the right representation.
Regina Barzilay (54:17.500)
And, you know, they couldn't get very far
Lex Fridman (54:22.240)
with this understanding while these models using,
Regina Barzilay (54:26.580)
you know, other sources actually capable
Lex Fridman (54:29.640)
to make a lot of progress.
Regina Barzilay (54:31.680)
Now, I'm not naive to think
Lex Fridman (54:33.960)
that we are in this paradise space in NLP.
Lex Fridman (54:36.780)
And sure as you know,
Lex Fridman (54:38.580)
that when we slightly change the domain
Lex Fridman (54:40.860)
and when we decrease the amount of training,
Lex Fridman (54:42.620)
it can do like really bizarre and funny thing.
Lex Fridman (54:44.740)
But I think it's just a matter
Lex Fridman (54:46.500)
of improving generalization capacity,
Regina Barzilay (54:48.540)
which is just a technical question.
Lex Fridman (54:51.500)
Wow, so that's the question.
Lex Fridman (54:54.340)
How much of language understanding can be solved
Lex Fridman (54:57.720)
with deep neural networks?
Regina Barzilay (54:59.180)
In your intuition, I mean, it's unknown, I suppose.
Lex Fridman (55:03.740)
But as we start to creep towards romantic notions
Regina Barzilay (55:07.660)
of the spirit of the Turing test
Lex Fridman (55:10.620)
and conversation and dialogue
Lex Fridman (55:14.220)
and something that maybe to me or to us,
Lex Fridman (55:18.340)
so the humans feels like it needs real understanding.
Lex Fridman (55:21.620)
How much can that be achieved
Lex Fridman (55:23.500)
with these neural networks or statistical methods?
Lex Fridman (55:27.180)
So I guess I am very much driven by the outcomes.
Lex Fridman (55:33.340)
Can we achieve the performance
Lex Fridman (55:35.420)
which would be satisfactory for us for different tasks?
Lex Fridman (55:40.700)
Now, if you again look at machine translation system,
Regina Barzilay (55:43.020)
which are trained on large amounts of data,
Lex Fridman (55:46.020)
they really can do a remarkable job
Regina Barzilay (55:48.780)
relatively to where they've been a few years ago.
Lex Fridman (55:51.300)
And if you project into the future,
Regina Barzilay (55:54.620)
if it will be the same speed of improvement, you know,
Lex Fridman (55:59.380)
this is great.
Regina Barzilay (56:00.220)
Now, does it bother me
Lex Fridman (56:01.060)
that it's not doing the same translation as we are doing?
Regina Barzilay (56:04.860)
Now, if you go to cognitive science,
Lex Fridman (56:06.620)
we still don't really understand what we are doing.
Regina Barzilay (56:10.460)
I mean, there are a lot of theories
Lex Fridman (56:11.860)
and there's obviously a lot of progress and studying,
Lex Fridman (56:13.840)
but our understanding what exactly goes on in our brains
Lex Fridman (56:17.540)
when we process language is still not crystal clear
Lex Fridman (56:21.020)
and precise that we can translate it into machines.
Lex Fridman (56:25.460)
What does bother me is that, you know,
Regina Barzilay (56:29.220)
again, that machines can be extremely brittle
Lex Fridman (56:31.700)
when you go out of your comfort zone
Regina Barzilay (56:33.980)
of when there is a distributional shift
Lex Fridman (56:36.060)
between training and testing.
Lex Fridman (56:37.300)
And it have been years and years,
Lex Fridman (56:39.020)
every year when I teach an LP class,
Regina Barzilay (56:41.320)
now show them some examples of translation
Lex Fridman (56:43.560)
from some newspaper in Hebrew or whatever, it was perfect.
Lex Fridman (56:47.300)
And then I have a recipe that Tomi Yakel's system
Lex Fridman (56:51.300)
sent me a while ago and it was written in Finnish
Regina Barzilay (56:53.900)
of Karelian pies.
Lex Fridman (56:55.720)
And it's just a terrible translation.
Regina Barzilay (56:59.280)
You cannot understand anything what it does.
Lex Fridman (57:01.460)
It's not like some syntactic mistakes, it's just terrible.
Lex Fridman (57:04.180)
And year after year, I tried and will translate
Lex Fridman (57:07.020)
and year after year, it does this terrible work
Regina Barzilay (57:08.980)
because I guess, you know, the recipes
Lex Fridman (57:10.980)
are not a big part of their training repertoire.
Regina Barzilay (57:14.580)
So, but in terms of outcomes, that's a really clean,
Lex Fridman (57:19.020)
good way to look at it.
Regina Barzilay (57:21.100)
I guess the question I was asking is,
Lex Fridman (57:24.060)
do you think, imagine a future,
Lex Fridman (57:27.700)
do you think the current approaches can pass
Lex Fridman (57:30.540)
the Turing test in the way,
Lex Fridman (57:34.700)
in the best possible formulation of the Turing test?
Lex Fridman (57:37.060)
Which is, would you wanna have a conversation
Lex Fridman (57:39.460)
with a neural network for an hour?
Lex Fridman (57:42.340)
Oh God, no, no, there are not that many people
Regina Barzilay (57:45.820)
that I would want to talk for an hour, but.
Lex Fridman (57:48.380)
There are some people in this world, alive or not,
Regina Barzilay (57:51.500)
that you would like to talk to for an hour.
Lex Fridman (57:53.260)
Could a neural network achieve that outcome?
Lex Fridman (57:56.700)
So I think it would be really hard to create
Lex Fridman (57:58.860)
a successful training set, which would enable it
Regina Barzilay (58:02.300)
to have a conversation, a contextual conversation
Lex Fridman (58:04.980)
for an hour.
Lex Fridman (58:05.820)
Do you think it's a problem of data, perhaps?
Lex Fridman (58:08.140)
I think in some ways it's not a problem of data,
Regina Barzilay (58:09.940)
it's a problem both of data and the problem of
Lex Fridman (58:13.620)
the way we're training our systems,
Regina Barzilay (58:15.780)
their ability to truly, to generalize,
Lex Fridman (58:18.060)
to be very compositional.
Regina Barzilay (58:19.300)
In some ways it's limited in the current capacity,
Lex Fridman (58:23.220)
at least we can translate well,
Regina Barzilay (58:27.980)
we can find information well, we can extract information.
Lex Fridman (58:32.540)
So there are many capacities in which it's doing very well.
Lex Fridman (58:35.180)
And you can ask me, would you trust the machine
Lex Fridman (58:38.000)
to translate for you and use it as a source?
Regina Barzilay (58:39.820)
I would say absolutely, especially if we're talking about
Lex Fridman (58:42.580)
newspaper data or other data which is in the realm
Regina Barzilay (58:45.660)
of its own training set, I would say yes.
Lex Fridman (58:48.900)
But having conversations with the machine,
Regina Barzilay (58:52.900)
it's not something that I would choose to do.
Lex Fridman (58:56.460)
But I would tell you something, talking about Turing tests
Lex Fridman (58:59.420)
and about all this kind of ELISA conversations,
Lex Fridman (59:02.940)
I remember visiting Tencent in China
Lex Fridman (59:05.540)
and they have this chat board and they claim
Lex Fridman (59:07.620)
there is really humongous amount of the local population
Regina Barzilay (59:10.780)
which for hours talks to the chat board.
Lex Fridman (59:12.940)
To me it was, I cannot believe it,
Lex Fridman (59:15.340)
but apparently it's documented that there are some people
Lex Fridman (59:18.000)
who enjoy this conversation.
Lex Fridman (59:20.760)
And it brought to me another MIT story
Lex Fridman (59:24.540)
about ELISA and Weisenbaum.
Regina Barzilay (59:26.980)
I don't know if you're familiar with the story.
Lex Fridman (59:29.340)
So Weisenbaum was a professor at MIT
Lex Fridman (59:31.020)
and when he developed this ELISA,
Lex Fridman (59:32.580)
which was just doing string matching,
Regina Barzilay (59:34.620)
very trivial, like restating of what you said
Lex Fridman (59:38.540)
with very few rules, no syntax.
Regina Barzilay (59:41.260)
Apparently there were secretaries at MIT
Lex Fridman (59:43.740)
that would sit for hours and converse with this trivial thing
Lex Fridman (59:48.180)
and at the time there was no beautiful interfaces
Lex Fridman (59:50.180)
so you actually need to go through the pain
Regina Barzilay (59:51.820)
of communicating.
Lex Fridman (59:53.540)
And Weisenbaum himself was so horrified by this phenomenon
Regina Barzilay (59:56.940)
that people can believe enough to the machine
Lex Fridman (59:59.300)
that you just need to give them the hint
🔗 相关节目